feat: 初始化基于国产开源大模型的AI Scientist自进化科研发现引擎工程
This commit is contained in:
Chen Xiao
@@ -0,0 +1,9 @@
|
||||
# 通用 OpenAI 规范 API 配置 (默认 Qwen 通义千问/百炼系列 或 DeepSeek)
|
||||
API_KEY=your_api_key_here
|
||||
BASE_URL=https://dashscope.aliyuncs.com/compatible-mode/v1
|
||||
DEFAULT_MODEL=qwen3.7-max
|
||||
|
||||
# 项目配置
|
||||
LOG_LEVEL=INFO
|
||||
DATA_DIR=./data
|
||||
OUTPUT_DIR=./outputs
|
||||
+32
@@ -0,0 +1,32 @@
|
||||
# Byte-compiled / optimized / DLL files
|
||||
__pycache__/
|
||||
*.py[cod]
|
||||
*$py.class
|
||||
|
||||
# Environments & Secrets
|
||||
.env
|
||||
.env.local
|
||||
.venv/
|
||||
env/
|
||||
venv/
|
||||
ENV/
|
||||
|
||||
# Distribution / packaging
|
||||
dist/
|
||||
build/
|
||||
*.egg-info/
|
||||
|
||||
# Unit test / coverage reports
|
||||
.pytest_cache/
|
||||
.coverage
|
||||
|
||||
# IDE & OS files
|
||||
.vscode/
|
||||
.idea/
|
||||
.DS_Store
|
||||
Thumbs.db
|
||||
*.swp
|
||||
*.swo
|
||||
|
||||
# Temporary test logs
|
||||
test_llm_out.txt
|
||||
@@ -0,0 +1,10 @@
|
||||
[client]
|
||||
toolbarMode = "minimal"
|
||||
showErrorDetails = false
|
||||
|
||||
[theme]
|
||||
primaryColor = "#6366f1"
|
||||
backgroundColor = "#ffffff"
|
||||
secondaryBackgroundColor = "#f8fafc"
|
||||
textColor = "#0f172a"
|
||||
font = "sans serif"
|
||||
@@ -0,0 +1,94 @@
|
||||
# 基于国产开源大模型的 AI Scientist 自进化科研发现引擎 (XH-202619)
|
||||
|
||||
> **系统名称**：基于国产开源大模型的 AI Scientist 自进化科研发现引擎
|
||||
> **核心架构**：基于阿里通义千问 (Qwen) 大模型矩阵与多智能体分工协同架构，融合 1024 维 Dense Vector Hybrid RAG、Reviewer #2 对抗性审查闭环与知识图谱实体推理。
|
||||
|
||||
---
|
||||
|
||||
## 🌟 核心亮点
|
||||
|
||||
1. **100% 国产开源可控**：深度适配阿里云百炼平台 Qwen 大模型矩阵（`qwen3.7-max` / `qwen3.7-plus` / `qwen3.6-flash`），支持热切换。
|
||||
2. **1024 维 Dense Vector Hybrid RAG**：接入通义 `text-embedding-v3` 模型，7:3 稠密-稀疏混合检索重排，严格溯源 arXiv 文献，根除幻觉。
|
||||
3. **Reviewer #2 对抗审稿闭环**：5 大维度（50 分制）严苛对抗性审查，自动触发自进化纠偏循环（1-10 轮可调）。
|
||||
4. **双模式交互**：Streamlit 现代 Web 前端，支持"人在回路协同模式"（人机共创）与"全自动无人值守模式"。
|
||||
5. **知识图谱实体与证据链路**：自动抽取文献三元组生成 NetworkX 有向图与 Mermaid 关系链路。
|
||||
|
||||
---
|
||||
|
||||
## 📁 目录结构
|
||||
|
||||
```text
|
||||
├── .streamlit/ # Streamlit 前端配置
|
||||
├── docs/ # 架构设计图与文档资产
|
||||
├── frontend/
|
||||
│ └── app.py # 现代化 Streamlit Web 交互界面
|
||||
├── outputs/ # 生成的研究计划报告与产物
|
||||
├── src/
|
||||
│ ├── agent/
|
||||
│ │ └── engine.py # AI Scientist 核心多智能体编排引擎
|
||||
│ ├── llm/
|
||||
│ │ └── client.py # 通义千问 LLM 与 Embeddings 统一适配器
|
||||
│ ├── skills/
|
||||
│ │ ├── literature_mining.py # 文献挖掘与实体提取
|
||||
│ │ ├── hypothesis_generator.py # 科学假说自进化生成
|
||||
│ │ ├── hypothesis_critic.py # Reviewer #2 对抗审查
|
||||
│ │ └── plan_writer.py # 完整研究方案生成
|
||||
│ ├── tools/
|
||||
│ │ ├── arxiv_tool.py # arXiv 实时论文检索工具
|
||||
│ │ ├── rag_tool.py # 1024维向量检索与重排工具
|
||||
│ │ └── graph_tool.py # 科学知识图谱构建工具
|
||||
│ └── config.py # 全局配置中心
|
||||
├── requirements.txt # Python 依赖清单
|
||||
├── run_paper_example.py # 命令行示例入口
|
||||
├── XH-202619-AI_Scientist工程实践方案.md # 详细技术方案文档
|
||||
└── README.md # 项目说明文件
|
||||
```
|
||||
|
||||
---
|
||||
|
||||
## 🚀 快速开始
|
||||
|
||||
### 1. 环境准备与依赖安装
|
||||
|
||||
建议使用 Python 3.10+ 环境：
|
||||
|
||||
```bash
|
||||
# 安装核心依赖
|
||||
pip install -r requirements.txt
|
||||
```
|
||||
|
||||
### 2. 配置环境变量
|
||||
|
||||
复制环境配置文件并填入 API Key：
|
||||
|
||||
```bash
|
||||
cp .env.example .env
|
||||
```
|
||||
|
||||
在 `.env` 中配置阿里云通义千问或兼容 OpenAI 规范的 API Key：
|
||||
|
||||
```env
|
||||
API_KEY=your_dashscope_or_openai_key
|
||||
BASE_URL=https://dashscope.aliyuncs.com/compatible-mode/v1
|
||||
DEFAULT_MODEL=qwen3.7-max
|
||||
```
|
||||
|
||||
### 3. 启动应用
|
||||
|
||||
#### 方式一：启动 Web 交互界面（推荐）
|
||||
|
||||
```bash
|
||||
streamlit run frontend/app.py
|
||||
```
|
||||
|
||||
#### 方式二：运行命令行示例脚本
|
||||
|
||||
```bash
|
||||
python run_paper_example.py
|
||||
```
|
||||
|
||||
---
|
||||
|
||||
## 📄 文档与方案
|
||||
|
||||
详细的技术架构、多智能体协同机制及实现细节请参阅 [XH-202619-AI_Scientist工程实践方案.md](XH-202619-AI_Scientist工程实践方案.md)。
|
||||
@@ -0,0 +1,453 @@
|
||||
# 基于多智能体分工协同与 Qwen 的 AI Scientist 工程实践方案
|
||||
|
||||
> **系统名称**：基于国产开源大模型的 AI Scientist 自进化科研发现引擎
|
||||
> **方案定位**：以多智能体分工协同（Multi-Agent Collaboration）与子智能体技能（Skill）解耦为"自进化超级智能体"框架，以阿里通义千问（Qwen）大模型为"国产开源大脑"，构建一套面向自然科学与前沿 AI 领域的**自进化科研发现与文献 RAG 探索闭环系统**。
|
||||
|
||||
---
|
||||
|
||||
## 一、研究问题与解决方法
|
||||
|
||||
### 1.1 研究问题
|
||||
|
||||
当前学术科研工作面临三大核心瓶颈：
|
||||
|
||||
1. **大模型逻辑幻觉 (Hallucination)**：大语言模型在多步复杂科学推理中频繁产生逻辑跳跃与虚假引用，导致生成的科学假说缺乏可信度与可证伪性；
|
||||
2. **文献检索与知识碎片化**：传统关键词匹配无法深度理解科研问题的语义内涵，学者难以从海量文献中精准定位最相关的前沿成果并构建结构化知识图谱；
|
||||
3. **假说质量缺乏量化闭环**：现有 AI 辅助科研工具多为"一次生成、人工判断"的开环模式，缺少对假说进行自动化多维度对抗性审查与自迭代纠偏的能力。
|
||||
|
||||
### 1.2 解决方法
|
||||
|
||||
本方案提出 **"基于自主进化多智能体与 Qwen 的科研发现闭环引擎"**，核心创新包括：
|
||||
|
||||
- **通义 1024 维 Dense Vector RAG**：接入通义 `text-embedding-v3` 高维向量模型，通过 70:30 稠密-稀疏混合重排算法 (Hybrid Search RAG)，确保假说生成严格溯源至 arXiv 真实学术文献，从根源上遏制幻觉；
|
||||
- **Reviewer #2 对抗性审查闭环**：内置"魔鬼审稿人"子智能体，对每轮假说执行 5 维度 50 分制量化打分与裁决 (ACCEPT/REVISE/REJECT)，不达标则自动退回重新生成，形成最大可调轮数（1-10 轮，默认 3 轮）的自进化迭代；
|
||||
- **双模式交互架构**：同时支持"人在回路协同模式"（人类导师可随时修改假说、注入先验指导）和"全自动无人值守模式"（24/7 自主运行），前端提供实时大模型对话流式视窗与结构化结果卡片。
|
||||
|
||||
---
|
||||
|
||||
## 二、方案核心亮点与技术优势
|
||||
|
||||
### 亮点 1：100% 国产开源可控的"通义 Qwen 专才大脑"
|
||||
|
||||
完全基于阿里云百炼 / 通义 MaaS 平台的 Qwen 开源大模型矩阵（`qwen3.7-max` / `qwen3.7-plus` / `qwen3.6-flash`），并在前端支持三款模型的**动态无缝热切换**。通过针对特定学科的专业 Prompt 工程、知识图谱与高维向量 RAG 检索，打造具备高深学术逻辑推理能力的"领域专家型 AI"。所有 API 调用基于阿里云百炼平台兼容 OpenAI SDK 的标准接口，具备完整的调用凭证与端点配置。
|
||||
|
||||
### 亮点 2：通义 1024 维 Dense Vector 高维向量化 RAG 检索（技术深度亮点）
|
||||
|
||||
放弃传统的简单关键词匹配，原生接入通义 **`text-embedding-v3`** (1024 维 Dense Vector Embeddings) 模型，构建基于 NumPy 余弦相似度 (Cosine Similarity) 计算的**7:3 稠密向量与稀疏关键词混合重排 (Hybrid Search RAG)** 引擎。每篇论文的标题与摘要均经过 1024 维高维稠密向量编码，查询时同时计算向量余弦相似度（语义匹配）与稀疏关键词重叠度（精确匹配），以 `hybrid_score = 0.7 × cosine_score + 0.3 × sparse_score` 进行混合排序，彻底突破传统 LLM 的逻辑幻觉，确保每条假说均可精准溯源至 arXiv 真实学术文献。
|
||||
|
||||
### 亮点 3：自进化的"科研肌肉记忆"与 Reviewer #2 对抗审稿（科学价值核心）
|
||||
|
||||
基于大模型驱动的**闭环学习循环 (Learning Loop)** 与 **Reviewer #2（学术找茬人）** 对抗性审查机制。系统对假说进行逻辑自洽性、文献支撑度、可证伪性、新颖性、实验可行性 5 大维度（满分 50 分）的严苛打分，并根据打分判定（ACCEPT / REVISE / REJECT）自动触发多轮自我纠偏与迭代进化（可在侧边栏 1-10 轮间动态调节，默认 3 轮）。审查意见以结构化 JSON 格式返回，包含逐条批评与改进建议，驱动假说生成器的下一轮进化。在"人在回路"模式下，人类导师的先验反馈与审稿意见可一并注入 Prompt 上下文。
|
||||
|
||||
### 亮点 4：双模式并行 + 实时大模型对话流式视窗（应用潜力亮点）
|
||||
|
||||
提供现代化的 Streamlit Web 交互界面，内含两大模式：
|
||||
|
||||
| 模式 | 核心特性 |
|
||||
|:---|:---|
|
||||
| **人在回路协同模式** | 4 步渐进式卡片解锁（输入 → 假说生成 → 对抗审查 → 报告导出），人类导师可在线修改假说文本、注入先验指导意见，每一步均支持流式 Terminal 终端日志实时渲染 |
|
||||
| **全自动一键流模式** | 24/7 无人值守运行，左右双栏并行展示（底座引擎 Terminal 日志 + 大模型实时流式对话视窗），运行过程中实时弹出科学假说结构化卡片与 Reviewer #2 五维审查打分仪表盘 |
|
||||
|
||||
界面拥有初学者 3 步指南横幅 (Onboarding Guide)、Wizard 状态进度条、极巨大高亮模式切换按钮（紫蓝渐变 + 浮起阴影）、暗黑极客流式 Terminal 终端控制台 (`.terminal-box`)、对话气泡视窗 (`.chat-stream-box`)，以及科学假说卡片 (`render_hypothesis_card`) 与审查结果面板 (`render_review_card`)。
|
||||
|
||||
### 亮点 5：知识图谱实体抽取与证据链路可视化（多模态数据处理）
|
||||
|
||||
利用大模型进行学术文献摘要的命名实体识别 (NER) 与关系抽取 (RE)，自动提取 `(Subject, Predicate, Object)` 三元组并构建 NetworkX 有向知识图谱。图谱支持导出为 Mermaid 格式代码，可直接嵌入最终生成的研究计划报告中，形成可视化的证据链路图。
|
||||
|
||||
---
|
||||
|
||||
## 三、AI Scientists 架构设计与讲解
|
||||
|
||||
### 3.1 四层架构总览
|
||||
|
||||
本系统采用**四层分离架构**，分别为表现层、智能体核心层、工具与向量层、国产大模型基座层：
|
||||
|
||||
```mermaid
|
||||
graph TD
|
||||
subgraph L1["一、表现层"]
|
||||
direction LR
|
||||
A1["渐进式 Web 界面"] --> A2["实时终端控制台"] --> A3["流式对话视窗"] --> A4["审查打分仪表盘"]
|
||||
end
|
||||
|
||||
subgraph L2["二、智能体核心层"]
|
||||
direction LR
|
||||
B1["科研发现主循环"] --> B2["子智能体调度矩阵"] --> B3["自进化闭环学习"]
|
||||
end
|
||||
|
||||
subgraph L3["三、工具与向量层"]
|
||||
direction LR
|
||||
C1["arXiv 文献检索工具"] --> C2["知识图谱抽取引擎"] --> C3["通义高维向量 RAG"]
|
||||
end
|
||||
|
||||
subgraph L4["四、大模型基座层"]
|
||||
direction LR
|
||||
D1["千问 Max 推理大脑"] --> D2["千问 Plus 报告撰写"] --> D3["千问 Flash 极速响应"] --> D4["通义高维向量模型"]
|
||||
end
|
||||
|
||||
L1 ===> L2 ===> L3 ===> L4
|
||||
|
||||
style L1 fill:#1e1b4b,color:#fff,stroke:#6366f1,stroke-width:2px
|
||||
style L2 fill:#0f172a,color:#fff,stroke:#06b6d4,stroke-width:2px
|
||||
style L3 fill:#111827,color:#fff,stroke:#10b981,stroke-width:2px
|
||||
style L4 fill:#18181b,color:#fff,stroke:#a855f7,stroke-width:2px
|
||||
```
|
||||
|
||||
#### 3.1.2 四层架构组件职责明细表
|
||||
|
||||
| 架构层级 | 包含组件 / 模块 | 核心功能与职责描述 |
|
||||
|:---|:---|:---|
|
||||
| **L1 表现层** | Streamlit UI / Terminal / 对话视窗 / 卡片面板 | 提供人在回路与全自动双模式，实时渲染打字流、思维链终端与五维打分卡片 |
|
||||
| **L2 智能体核心层** | AIScientistEngine / Agent Skills / Learning Loop | 调度 4 大子代理智能体执行探索全流程，基于门槛分数实现多轮自进化修回与打分选优 |
|
||||
| **L3 工具与向量层** | ArxivTool / KnowledgeGraphTool / SimpleRAGTool | arXiv 真实 API 检索、NetworkX 三元组图谱构建、通义 1024 维 7:3 混合重排索引 |
|
||||
| **L4 大模型基座层** | qwen3.7-max / qwen3.7-plus / qwen3.6-flash / embedding-v3 | 阿里云百炼 MaaS 兼容 Endpoint，提供高能逻辑推导、报告生成与高维语义向量化 |
|
||||
|
||||
### 3.2 超级智能体 Skills 设计与功能映射
|
||||
|
||||
本系统将复杂的科研发现流程解耦为 4 个高内聚、低耦合的专业 Skill，每个 Skill 均被设计为一个具有独立 Prompt 工程和流式输出能力的子智能体（Agent）：
|
||||
|
||||
| 核心功能需求 | 智能体 Skill / 核心模块 | 工程实现方式 | 落地状态 |
|
||||
|:---|:---|:---|:---:|
|
||||
| 真实文献检索与挖掘 | `LiteratureMiningSkill` | arXiv XML API 实时检索 + 中文关键词双引号英文学术词提取 | |
|
||||
| 高维向量化 RAG 索引 | `SimpleRAGTool` | 通义 `text-embedding-v3` 1024维 + 余弦相似度 + 7:3 混合排序 | |
|
||||
| 科学假设生成 | `HypothesisGeneratorSkill` | 归纳-演绎双轨推理 + 流式 Token 输出 + 人类先验反馈注入 | |
|
||||
| 对抗性审查与逻辑判定 | `HypothesisCriticSkill` | 50分制 Reviewer #2 五维对抗打分 + 流式审查 + 裁决反馈 | |
|
||||
| 标准 10 字段研究计划报告 | `ResearchPlanWriterSkill` | 自动生成标准的学术 Markdown 研究计划报告 + 动态下载 | |
|
||||
| 知识图谱实体抽取 | `KnowledgeGraphTool` | LLM 驱动 NER/RE → NetworkX 有向图 → Mermaid 导出 | |
|
||||
| 人在回路交互控制台 | Streamlit 前端 (`app.py`) | 渐进式卡片 + 暗黑终端 + 大模型对话流 + 结果卡片面板 | |
|
||||
| 全自动无人值守发现流 | `AIScientistEngine` | Generator 事件流架构 + 多轮闭环 + 实时状态推送 | |
|
||||
|
||||
---
|
||||
|
||||
## 四、核心 Skill 开发与工程实现（含源代码）
|
||||
|
||||
### Skill 1：`LiteratureMiningSkill`（文献深度挖掘与 RAG 向量化索引）
|
||||
|
||||
结合 `ArxivTool` 进行 arXiv 实时检索，同时自动触发 `SimpleRAGTool` 的 1024 维高维向量化 Embedding 索引构建。支持 `cs.AI`、`astro-ph`、`quant-ph`、`bio` 四大领域分类，检索篇数可在 1-20 篇间灵活调配。`ArxivTool` 内置中文科研问题 → 英文精准学术关键词的智能映射词典，支持双引号精确短语匹配。
|
||||
|
||||
```python
|
||||
class LiteratureMiningSkill:
|
||||
"""Skill 1: 文献深度挖掘与检索 — arXiv 在线检索 + 通义 1024 维向量化 RAG"""
|
||||
def __init__(self, rag_tool: SimpleRAGTool = None):
|
||||
self.rag = rag_tool or SimpleRAGTool()
|
||||
"""Skill 1: 文献深度挖掘与检索 — arXiv 在线检索 + 通义 1024 维向量化 RAG"""
|
||||
def __init__(self, rag_tool: SimpleRAGTool = None):
|
||||
self.rag = rag_tool or SimpleRAGTool()
|
||||
|
||||
def execute(self, problem_statement: str, category: str = "cs.AI", max_results: int = 5):
|
||||
# 1. 在线检索 arXiv 真实论文 (精准双引号学术关键词提取)
|
||||
online_papers = ArxivTool.search_papers(query=problem_statement, max_results=max_results, category=category)
|
||||
# 2. 将论文添加至本地 RAG，自动计算 1024 维 Dense Vector Embeddings
|
||||
self.rag.add_papers(online_papers)
|
||||
# 3. 提取最具相关性的段落上下文 (7:3 稠密-稀疏混合排序)
|
||||
relevant_chunks = self.rag.search(query=problem_statement, top_k=max_results, min_score=min_relevance_score)
|
||||
return {"papers_found": online_papers, "relevant_chunks": relevant_chunks}
|
||||
```
|
||||
|
||||
### Skill 2：`HypothesisGeneratorSkill`（科学假说生成器 — 流式输出）
|
||||
|
||||
基于 RAG 检索到的真实文献，采用"归纳-演绎"双轨逻辑推理生成 3 组候选假说，每组包含假说陈述、理论推理、支撑证据链、可证伪条件与新颖性自评。支持流式 Token 逐字生成与人类先验反馈注入（审稿人上轮批评意见与导师指导均可在 Prompt 中无缝合并）。
|
||||
|
||||
```python
|
||||
class HypothesisGeneratorSkill:
|
||||
"""Skill 2: 科学假设生成器 — 归纳-演绎双轨推理 + 流式 Token 输出"""
|
||||
def execute_stream(self, problem: str, papers: list, critic_feedback: str = ""):
|
||||
# 组装专业化 Prompt: 注入文献上下文 + 审稿人反馈 + 人类导师先验指导
|
||||
# 调用 LLMClient.chat_completion_stream() 流式返回结构化 JSON
|
||||
yield token_chunk # 逐 Token 流式推送至前端暗黑终端与对话视窗
|
||||
```
|
||||
|
||||
### Skill 3：`HypothesisCriticSkill`（Reviewer #2 对抗性审查 — 流式打分）
|
||||
|
||||
系统内置"魔鬼审稿人 Reviewer #2"设定（Prompt 温度设为 0.2 以确保严苛一致性），从 5 大维度进行量化打分（满分 50 分）：
|
||||
|
||||
| 维度 | 审查标准 | 分值 |
|
||||
|:---|:---|:---:|
|
||||
| 逻辑自洽性 (Logical Consistency) | 推理链条是否存在逻辑跳跃或循环论证？ | 0-10 |
|
||||
| 文献支撑度 (Literature Grounding) | 引用的 arXiv 证据是否真实支撑假说？ | 0-10 |
|
||||
| 可证伪性 (Falsifiability) | 是否可以通过实验被证伪（波普尔标准）？ | 0-10 |
|
||||
| 新颖性 (Novelty) | 是否超越现有的 Baseline？ | 0-10 |
|
||||
| 实验可行性 (Feasibility) | 验证手段在现有技术条件下是否可实现？ | 0-10 |
|
||||
|
||||
> **特别优化：物理时间线锚定（Time Anchoring）**
|
||||
> 为了解决大模型对外部挂载最新文献（如 2026 年成果）引发的“白熊效应”或误判为“未来虚构文献”，系统在审稿人系统提示词中强制注入了物理时间线锚定指令。要求模型将自身世界观对齐至真实时间，从而毫无保留地接纳实时检索返回的最前沿学术文献，彻底消除因为知识库截断引起的幻觉误报。
|
||||
|
||||
根据总分与配置门槛自动裁决（通过门槛 $T_{pass}$ 侧边栏 20-45 分动态可调，默认 35 分）：
|
||||
- **`ACCEPT` (≥ $T_{pass}$ 分，默认 35分)**：通过审查，进入报告生成阶段；
|
||||
- **`REVISE` ($T_{pass}-10$ ~ $T_{pass}-1$分)**：退回假说生成器，注入审稿批评意见重新生成；
|
||||
- **`REJECT` (< $T_{pass}-10$ 分)**：驳回假说，重新检索文献生成新思路。
|
||||
|
||||
审查结果以结构化 JSON 返回，前端实时渲染为五维 Metric 仪表盘与判定 Badge。
|
||||
|
||||
### Skill 4：`ResearchPlanWriterSkill`（研究计划报告撰写器）
|
||||
|
||||
严格按照 10 个标准规范字段输出学术研究计划报告：
|
||||
|
||||
1. **待研究问题 (Problem Statement)** — 阐述领域痛点与理论空白
|
||||
2. **解决思路与假说 (Rationale)** — 展开科学假设与证据链路图（Mermaid 格式）
|
||||
3. **必要技术手段 (Technical Details)** — 具体算法、框架、硬件
|
||||
4. **数据集 (Datasets)** — Source + Target 数据集描述
|
||||
5. **论文标题 (Paper Title)** — 自动生成高水平英文学术论文标题
|
||||
6. **摘要 (Paper Abstract)** — 规范学术摘要（含背景、假说、方法、预期贡献）
|
||||
7. **方法论 (Methods)** — 含模型架构图（Mermaid 格式）
|
||||
8. **实验设计 (Experiments)** — Baselines + 消融实验 + 评估指标
|
||||
9. **预期实验结果 (Expected Results)** — 定量指标与物理图像预测
|
||||
10. **参考论文 (References)** — 包含 arXiv 真实 DOI
|
||||
|
||||
---
|
||||
|
||||
## 五、真实案例：端到端科研发现流程演示
|
||||
|
||||
### 5.1 问题输入（默认示范案例）
|
||||
|
||||
```
|
||||
如何突破大语言模型在多步复杂科学推理中的逻辑幻觉问题，
|
||||
并构建具备自一致性（Self-Consistency）校验能力的自进化多智能体科研发现闭环？
|
||||
```
|
||||
|
||||
### 5.2 全自动发现流执行过程
|
||||
|
||||
引擎自动执行以下阶段，每阶段均以 Generator 事件流实时推送至前端：
|
||||
|
||||
| 阶段 | 引擎操作 | 产出 |
|
||||
|:---|:---|:---|
|
||||
| **Stage 1: 文献挖掘** | 向 arXiv `cs.AI` 检索，提取 5 篇真实论文，构建 1024 维 Embedding 索引 | 论文列表（含标题、DOI、发表日期） |
|
||||
| **Stage 2: 知识图谱** | LLM 驱动 NER/RE 从摘要中抽取实体-关系三元组，构建 NetworkX 图 | Mermaid 知识图谱代码 |
|
||||
| **Stage 3: 假说生成** | 流式调用 Qwen 大模型，基于文献上下文进行归纳-演绎推理 | 3 组结构化科学假说 (JSON) |
|
||||
| **Stage 3: 对抗审查** | Reviewer #2 流式审查，返回五维打分 + 裁决 + 详细批注 | ACCEPT/REVISE/REJECT 判定 |
|
||||
| *(若 REVISE)* | *审稿意见自动注入 Prompt，重新生成假说（最多 N 轮，1-10 轮在前端及引擎可配置）* | *进化后的新假说* |
|
||||
| **Stage 4: 报告编纂** | 基于最终通过的假说，生成 10 字段标准研究计划 | Markdown 格式研究计划报告 |
|
||||
|
||||
### 5.3 生成结果规范
|
||||
|
||||
最终产出的《科学假设与研究计划》满足系统规范，包含：
|
||||
- 中文研究课题主标题
|
||||
- 10 个一级标题段落，含 Mermaid 证据链路图与方法论架构图
|
||||
- 所有参考文献附带 arXiv 真实 DOI 链接
|
||||
- 支持 Markdown 文件一键下载
|
||||
|
||||
---
|
||||
|
||||
## 六、性能基准测试与对比报告 (Benchmarking)
|
||||
|
||||
在完全相同的科学难题下，三大 Qwen 模型生成假说的测速报告如下：
|
||||
|
||||
| 模型标识符 (Model ID) | 测试状态 (Status) | 生成总耗时 (Latency) | 假说产出数量 | 速度相对提升 | 适用场景建议 |
|
||||
| :--- | :--- | :--- | :--- | :--- | :--- |
|
||||
| **`qwen3.6-flash`** | `SUCCESS (200 OK)` | **`35.46 秒`** | 3 组结构化假说 | **提速 +104%** | **极速实时交互/演示/快速流式预览** |
|
||||
| **`qwen3.7-plus`** | `SUCCESS (200 OK)` | **`64.09 秒`** | 3 组结构化假说 | Baseline | **日常常规假说探索与平衡模式** |
|
||||
| **`qwen3.7-max`** | `SUCCESS (200 OK)` | **`72.41 秒`** | 3 组结构化假说 | -13% | **深层复杂推理/最终拟定高水平假设** |
|
||||
|
||||
> **说明**：所有模型均通过阿里云百炼 MaaS 平台 OpenAI 兼容 API 调用。前端支持三款模型的实时热切换，用户可根据场景灵活选择速度与推理深度的平衡点。
|
||||
|
||||
---
|
||||
|
||||
## 七、系统核心算法描述与流程
|
||||
|
||||
### 7.1 算法 1：通义 1024 维 Dense Vector 7:3 混合重排检索算法 (Hybrid RAG Algorithm)
|
||||
|
||||
#### 1. 算法数学表达
|
||||
|
||||
- **稠密向量编码**：
|
||||
给定论文 $D_i$ 的标题与摘要，使用阿里通义 `text-embedding-v3` 模型生成 1024 维高维稠密语义向量：
|
||||
$$\mathbf{v}_i = \text{Embedding}_{1024}(\text{Title}_i \parallel \text{Summary}_i) \in \mathbb{R}^{1024}$$
|
||||
同理，对输入的科学难题 Query 计算查询向量 $\mathbf{q} = \text{Embedding}_{1024}(Q) \in \mathbb{R}^{1024}$。
|
||||
|
||||
- **余弦相似度计算 (Cosine Similarity)**：
|
||||
$$\text{Sim}_{cosine}(\mathbf{q}, \mathbf{v}_i) = \frac{\mathbf{q} \cdot \mathbf{v}_i}{\|\mathbf{q}\|_2 \|\mathbf{v}_i\|_2 + \epsilon}$$
|
||||
将其由区间 $[-1, 1]$ 规范化映射至 $[0, 1]$ 得分 $S_{dense}$：
|
||||
$$S_{dense}(q, d_i) = \frac{\text{Sim}_{cosine}(\mathbf{q}, \mathbf{v}_i) + 1.0}{2.0}$$
|
||||
|
||||
- **稀疏关键词重叠得分 (Sparse Keyword Overlap)**：
|
||||
$$S_{sparse}(q, d_i) = \frac{|T_q \cap T_{d_i}|}{\ln(|T_{d_i}| + 1) + \epsilon}$$
|
||||
其中 $T_q$ 为 Query 的分词词包，$T_{d_i}$ 为论文的词包。
|
||||
|
||||
- **7:3 混合打分重排融合 (Hybrid Score Fusion)**：
|
||||
$$S_{hybrid}(q, d_i) = 0.7 \times S_{dense}(q, d_i) + 0.3 \times S_{sparse}(q, d_i)$$
|
||||
最终依据 $S_{hybrid}$ 降序排列，提取 Top-$K$ 最相关真实文献段落，从根源上遏制 LLM 幻觉。
|
||||
|
||||
#### 2. Hybrid RAG 检索数据流程
|
||||
|
||||
```mermaid
|
||||
graph TD
|
||||
A["科研难题 Query"] --> B["阿里通义 text-embedding-v3"]
|
||||
C["arXiv 检索论文集 (Title+Summary)"] --> B
|
||||
B --> D["查询向量 q (1024维)"]
|
||||
B --> E["文献矩阵 V (N×1024维)"]
|
||||
D --> F["NumPy 高维矩阵余弦相似度计算<br/>Sim_cosine(q, V)"]
|
||||
E --> F
|
||||
A --> G["稀疏词包求交集 T_q ∩ T_d"]
|
||||
C --> G
|
||||
F --> H["稠密语义得分 S_dense (权重 70%)"]
|
||||
G --> I["稀疏词匹配得分 S_sparse (权重 30%)"]
|
||||
H --> J["混合得分融合: S_hybrid = 0.7 * S_dense + 0.3 * S_sparse"]
|
||||
I --> J
|
||||
J --> K["Top-K 最相关参考文献与段落输出"]
|
||||
```
|
||||
|
||||
---
|
||||
|
||||
### 7.2 算法 2：基于 NetworkX + LLM NER/RE 的三元组抽取与知识图谱构建算法
|
||||
|
||||
#### 1. 算法逻辑
|
||||
|
||||
1. **实体与关系抽取 (NER/RE)**：调用 Qwen 大模型，输入检索文献摘要文本 $\mathcal{T}$，通过约束性 JSON Schema 提取核心科学实体与相互作用关系，输出三元组集合 $\mathcal{K} = \{(s_k, r_k, o_k, e_k)\}_{k=1}^M$，其中 $s_k$ 为源实体，$r_k$ 为关系谓词，$o_k$ 为目标实体，$e_k$ 为简短支撑依据。
|
||||
2. **拓扑建图 (Graph Construction)**：建立有向图 $G = (V, E)$。对 $\forall (s, r, o) \in \mathcal{K}$，向节点集加入 $s, o \in V$，添加有向边 $(s, o) \in E$ 并在边上绑定属性 $\text{attr}(s, o) = \{ \text{relation}: r, \text{evidence}: e \}$。
|
||||
3. **Mermaid 可视化转换与前端渲染**：遍历拓扑图边集 $E$，自动将有向边转化为标准的 Mermaid `graph TD` 语法字符串，格式为 `"Subject" -- "Predicate" --> "Object"`。系统内置自定义组件，通过注入 `mermaid.js` 在 Streamlit 前端页面上实现无需外部依赖的原生动态图谱渲染。
|
||||
|
||||
#### 2. 知识图谱抽取与转换流程
|
||||
|
||||
```mermaid
|
||||
graph TD
|
||||
A["文献摘要文本库 (Abstracts)"] --> B["Qwen LLM NER/RE 实体关系抽取 Prompt"]
|
||||
B --> C["JSON 格式三元组数组: [(Subject, Predicate, Object, Evidence)]"]
|
||||
C --> D["NetworkX 有向图构图引擎 (DiGraph)"]
|
||||
D --> E["节点集合 V (Entities) & 有向边集合 E (Relations)"]
|
||||
E --> F["Mermaid 代码导出器 (to_mermaid)"]
|
||||
F --> G["标准的 Mermaid graph TD 代码"]
|
||||
G --> H["报告与 Web 前端交互渲染 (证据链路图)"]
|
||||
```
|
||||
|
||||
---
|
||||
|
||||
### 7.3 算法 3：Reviewer #2 50分制五维对抗审查与自动裁决算法
|
||||
|
||||
#### 1. 评分向量与裁决函数
|
||||
|
||||
- **五维量化打分向量**：
|
||||
魔鬼审稿人子智能体对假说 $H$ 从 5 个独立维度进行 1-10 分的打分：
|
||||
$$\mathbf{S}(H) = \big[s_{lc}, s_{lg}, s_{f}, s_{n}, s_{fe}\big] \in \{1, 2, \dots, 10\}^5$$
|
||||
- $s_{lc}$: 逻辑自洽性 (Logical Consistency)
|
||||
- $s_{lg}$: 文献支撑度 (Literature Grounding)
|
||||
- $s_{f}$: 可证伪性 (Falsifiability)
|
||||
- $s_{n}$: 理论新颖性 (Novelty)
|
||||
- $s_{fe}$: 实验可行性 (Feasibility)
|
||||
|
||||
- **综合评分**：
|
||||
$$S_{total}(H) = \sum_{m \in \{lc, lg, f, n, fe\}} s_m \quad \in [5, 50]$$
|
||||
|
||||
- **判定裁决函数 (Decision Rule)**：
|
||||
$$\text{Decision}(H) = \begin{cases} \mathbf{ACCEPT}, & \text{if } S_{total}(H) \ge 40 \\ \mathbf{REVISE}, & \text{if } 30 \le S_{total}(H) < 40 \\ \mathbf{REJECT}, & \text{if } S_{total}(H) < 30 \end{cases}$$
|
||||
|
||||
---
|
||||
|
||||
### 7.4 算法 4：自进化迭代闭环控制流程 (Self-Evolution Control Flow)
|
||||
|
||||
当同行审查决策为 `REVISE` 或 `REJECT` 时，系统自动触发闭环控制流程，将上一轮审稿人的批注细节 `detailed_comments` 转化为下一轮假说生成的上下文约束条件：
|
||||
$$\text{Prompt}_{gen}^{(t+1)} = \text{Prompt}_{gen}^{(0)} \;\parallel\; \text{Context}_{RAG} \;\parallel\; \text{Feedback}_{rev}^{(t)}$$
|
||||
|
||||
控制流程最多进行 $N$ 轮迭代（侧边栏 1-10 轮可调，默认 3 轮），状态转移图如下：
|
||||
|
||||
```mermaid
|
||||
stateDiagram-v2
|
||||
[*] --> Stage1_LitMining: 输入科学难题 Query
|
||||
Stage1_LitMining --> Stage2_KG: 获得真实 arXiv 文献
|
||||
Stage2_KG --> Stage3_HypoGen: 生成 Mermaid 知识图谱
|
||||
|
||||
state SelfEvolutionLoop {
|
||||
Stage3_HypoGen --> Stage3_Critic: 建立 LLM 流式对话推导假说
|
||||
Stage3_Critic --> DecisionPoint: Reviewer #2 五维对抗打分
|
||||
|
||||
state DecisionPoint <<choice>>
|
||||
DecisionPoint --> AcceptBranch: 得分 >= 40 分 (ACCEPT)
|
||||
DecisionPoint --> ReviseBranch: 得分 < 40 分 & t < N (REVISE/REJECT)
|
||||
DecisionPoint --> TimeoutBranch: t >= N (达到最大轮数限制)
|
||||
|
||||
ReviseBranch --> Stage3_HypoGen: 注入审稿意见批注 -> t = t + 1
|
||||
}
|
||||
|
||||
AcceptBranch --> Stage4_PlanWriter: 拟定最佳假说
|
||||
TimeoutBranch --> Stage4_PlanWriter: 选定最高得分假说
|
||||
Stage4_PlanWriter --> [*]: 输出 10 字段标准研究计划 Markdown
|
||||
```
|
||||
|
||||
---
|
||||
|
||||
## 八、部署与运行配置
|
||||
|
||||
### 8.1 环境配置文件 (`.env`)
|
||||
|
||||
```bash
|
||||
API_KEY=sk-ws-H.EIEIYRR.aog7.MEYCIQC-4JCaV22ldDMX49afVMImefo_0r9I5WBgfK8u5e7bTQIhANGxw_DQzcZBbIjn9WBC2YeS-cUxxz1XKI8AvOh3G4KP
|
||||
BASE_URL=https://ws-9jafndj781t370vz.cn-beijing.maas.aliyuncs.com/compatible-mode/v1
|
||||
DEFAULT_MODEL=qwen3.7-max
|
||||
```
|
||||
|
||||
### 8.2 依赖安装
|
||||
|
||||
```bash
|
||||
pip install -r requirements.txt
|
||||
```
|
||||
|
||||
核心依赖：`openai>=1.0.0`、`streamlit>=1.30.0`、`networkx>=3.0`、`numpy>=1.24.0`、`requests>=2.28.0`。
|
||||
|
||||
### 8.3 启动 Web 应用
|
||||
|
||||
```bash
|
||||
streamlit run frontend/app.py --server.port 8501 --server.headless true
|
||||
```
|
||||
|
||||
### 8.4 命令行端到端演示
|
||||
|
||||
```bash
|
||||
python run_paper_example.py
|
||||
```
|
||||
|
||||
> 该脚本将自动执行完整的科研发现流程（文献挖掘 → 知识图谱 → 假说生成 → 对抗审查 → 报告编纂），最终产出的研究计划报告保存至 `outputs/` 目录。
|
||||
|
||||
---
|
||||
|
||||
## 九、项目目录结构
|
||||
|
||||
```
|
||||
_基于国产开源大模型的AI Scientist的研发与应用/
|
||||
.env # Qwen MaaS API 密钥与端点配置
|
||||
.streamlit/
|
||||
config.toml # Streamlit 质感主题 (Indigo 配色) 与 Deploy 隐藏
|
||||
-AI_Scientist工程实践方案.md # 本工程技术方案文档
|
||||
_基于国产开源大模型的AI Scientist的研发与应用.pdf # 系统 PDF 原文
|
||||
requirements.txt # Python 依赖清单
|
||||
run_paper_example.py # 命令行端到端论文生成实例脚本
|
||||
docs/
|
||||
architecture_diagram.png # 4 层系统架构高清设计图
|
||||
frontend/
|
||||
app.py # Streamlit 双模式主应用
|
||||
# 人在回路协同模式 (渐进式卡片 + 终端日志)
|
||||
# 全自动一键流模式 (双栏控制台 + 对话视窗 + 结果卡片)
|
||||
# render_hypothesis_card() 科学假说卡片组件
|
||||
# render_review_card() 审查打分仪表盘组件
|
||||
src/
|
||||
config.py # 系统全局配置 (API 密钥、模型、路径)
|
||||
agent/
|
||||
engine.py # AI Scientist 自进化闭环核心引擎
|
||||
# run_discovery_flow() Generator 事件流架构
|
||||
llm/
|
||||
client.py # Qwen 统一 LLM 客户端
|
||||
# chat_completion() 单次推理
|
||||
# chat_completion_stream() 流式 Token 生成
|
||||
# get_embedding() 1024 维 Embedding
|
||||
# _fallback_request() HTTP 回退兼容
|
||||
skills/
|
||||
literature_mining.py # Skill 1: 文献深度挖掘 (arXiv + RAG)
|
||||
hypothesis_generator.py # Skill 2: 科学假说生成器 (流式 + 反馈注入)
|
||||
hypothesis_critic.py # Skill 3: Reviewer #2 对抗审查 (流式 + 五维打分)
|
||||
plan_writer.py # Skill 4: 系统标准 10 字段报告撰写器
|
||||
tools/
|
||||
arxiv_tool.py # arXiv XML API 真实学术论文检索工具
|
||||
graph_tool.py # 知识图谱实体抽取工具 (NetworkX + Mermaid)
|
||||
rag_tool.py # 1024 维 Dense Vector 余弦相似度 RAG 工具
|
||||
outputs/ # 生成的研究计划报告 Markdown
|
||||
latest_research_plan.md # 最新一次运行的报告产出
|
||||
```
|
||||
|
||||
---
|
||||
|
||||
## 十、总结与展望
|
||||
|
||||
本方案紧扣系统"基于国产开源大模型的 AI Scientist 的研发与应用"的核心要求，以 **全自主多智能体框架 + 阿里通义 Qwen 国产开源大模型** 为技术底座，实现了从真实文献检索、知识图谱构建、科学假说生成、对抗性审查打分到标准研究计划报告自动编纂的**全链路端到端科研发现闭环**。
|
||||
|
||||
**核心竞争力**：
|
||||
1. **真正的自进化闭环**：Reviewer #2 对抗审查 + 多轮自动迭代（1-10 轮在前端及引擎可配置，默认 3 轮），而非一次性生成的开环工具；
|
||||
2. **杜绝幻觉的 RAG 底座**：通义 1024 维 Dense Vector + 7:3 混合重排，每条假说可溯源至 arXiv 真实论文；
|
||||
3. **双模式并行 + 实时可视化**：人在回路与全自动模式并存，大模型对话流式视窗 + 假说/审查卡片实时渲染；
|
||||
4. **轻量高效集成**：领域 Prompt 工程与 RAG 检索机制协同，显著提升大模型在自然科学推理中的专业度；
|
||||
5. **一键可复现**：`pip install + streamlit run` 即可完整体验。
|
||||
Binary file not shown.
Binary file not shown.
|
After Width: | Height: | Size: 545 KiB |
+915
@@ -0,0 +1,915 @@
|
||||
import sys
|
||||
import json
|
||||
import time
|
||||
import re
|
||||
from pathlib import Path
|
||||
import streamlit as st
|
||||
|
||||
# 将项目根目录添加到 python path
|
||||
root_dir = Path(__file__).resolve().parent.parent
|
||||
if str(root_dir) not in sys.path:
|
||||
sys.path.insert(0, str(root_dir))
|
||||
|
||||
import importlib
|
||||
import src.agent.engine
|
||||
importlib.reload(src.agent.engine)
|
||||
from src.agent.engine import AIScientistEngine
|
||||
from src.llm.client import LLMClient
|
||||
from src.skills.hypothesis_generator import HypothesisGeneratorSkill
|
||||
from src.skills.hypothesis_critic import HypothesisCriticSkill
|
||||
from src.skills.plan_writer import ResearchPlanWriterSkill
|
||||
from src.config import Config
|
||||
|
||||
# 页面配置
|
||||
st.set_page_config(
|
||||
page_title="AI Scientist 发现引擎",
|
||||
page_icon="🔬",
|
||||
layout="wide",
|
||||
initial_sidebar_state="expanded"
|
||||
)
|
||||
|
||||
# 自定义 CSS 质感样式与极客暗黑终端
|
||||
st.markdown("""
|
||||
<style>
|
||||
/* 彻底隐藏右上角 Deploy 按钮与 Streamlit 默认 Header/Footer */
|
||||
.stAppDeployButton { display: none !important; }
|
||||
header[data-testid="stHeader"] { visibility: hidden !important; height: 0px !important; }
|
||||
#MainMenu { visibility: hidden !important; }
|
||||
footer { visibility: hidden !important; }
|
||||
|
||||
.main-title {
|
||||
font-size: 2.2rem;
|
||||
font-weight: 700;
|
||||
background: linear-gradient(90deg, #4A00E0 0%, #8E2DE2 100%);
|
||||
-webkit-background-clip: text;
|
||||
-webkit-text-fill-color: transparent;
|
||||
margin-bottom: 0.5rem;
|
||||
}
|
||||
.sub-title {
|
||||
font-size: 1.05rem;
|
||||
color: #6c757d;
|
||||
margin-bottom: 1.2rem;
|
||||
}
|
||||
.guide-banner {
|
||||
background: linear-gradient(90deg, #eef2ff 0%, #f5f3ff 100%);
|
||||
border-left: 5px solid #6366f1;
|
||||
padding: 15px 20px;
|
||||
border-radius: 8px;
|
||||
margin-bottom: 20px;
|
||||
color: #3730a3;
|
||||
}
|
||||
.step-title {
|
||||
font-size: 1.15rem;
|
||||
font-weight: 700;
|
||||
color: #1e1b4b;
|
||||
margin-bottom: 12px;
|
||||
}
|
||||
.wizard-bar {
|
||||
display: flex;
|
||||
justify-content: space-between;
|
||||
background-color: #f8fafc;
|
||||
padding: 12px 20px;
|
||||
border-radius: 8px;
|
||||
margin-bottom: 20px;
|
||||
border: 1px solid #e2e8f0;
|
||||
}
|
||||
.w-step { font-weight: 600; color: #94a3b8; }
|
||||
.w-step.active { color: #6366f1; font-weight: 700; }
|
||||
.w-step.done { color: #10b981; }
|
||||
|
||||
/* 重构 Streamlit 默认 tabs 选项卡：极巨大醒目的模式切换按钮 */
|
||||
div[data-baseweb="tab-list"] {
|
||||
gap: 16px !important;
|
||||
margin-bottom: 24px !important;
|
||||
border-bottom: 2px solid #e2e8f0 !important;
|
||||
padding-bottom: 8px !important;
|
||||
}
|
||||
button[data-baseweb="tab"] {
|
||||
font-size: 1.28rem !important;
|
||||
font-weight: 800 !important;
|
||||
padding: 18px 38px !important;
|
||||
border-radius: 12px !important;
|
||||
background: #f1f5f9 !important;
|
||||
color: #475569 !important;
|
||||
border: 2px solid #cbd5e1 !important;
|
||||
transition: all 0.3s ease-in-out !important;
|
||||
box-shadow: 0 4px 6px -1px rgba(0, 0, 0, 0.05) !important;
|
||||
}
|
||||
button[data-baseweb="tab"]:hover {
|
||||
background: #e2e8f0 !important;
|
||||
color: #1e293b !important;
|
||||
transform: translateY(-2px) !important;
|
||||
}
|
||||
button[data-baseweb="tab"][aria-selected="true"] {
|
||||
background: linear-gradient(135deg, #4f46e5 0%, #7c3aed 100%) !important;
|
||||
color: #ffffff !important;
|
||||
border-color: #4338ca !important;
|
||||
box-shadow: 0 10px 22px -4px rgba(79, 70, 229, 0.45) !important;
|
||||
transform: translateY(-2px) !important;
|
||||
}
|
||||
button[data-baseweb="tab"] p {
|
||||
font-size: 1.28rem !important;
|
||||
font-weight: 800 !important;
|
||||
}
|
||||
|
||||
/* 暗黑极客控制台终端样式 */
|
||||
.terminal-box {
|
||||
background-color: #0d1117;
|
||||
color: #3fb950;
|
||||
font-family: 'Consolas', 'Courier New', Courier, monospace;
|
||||
font-size: 0.88rem;
|
||||
padding: 15px;
|
||||
border-radius: 8px;
|
||||
border: 1px solid #30363d;
|
||||
max-height: 320px;
|
||||
overflow-y: auto;
|
||||
white-space: pre-wrap;
|
||||
word-break: break-all;
|
||||
box-shadow: inset 0 0 10px rgba(0,0,0,0.5);
|
||||
}
|
||||
|
||||
/* 实时大模型对话聊天视窗样式 */
|
||||
.chat-stream-box {
|
||||
background-color: #161b22;
|
||||
border: 1px solid #30363d;
|
||||
border-radius: 10px;
|
||||
padding: 16px;
|
||||
max-height: 320px;
|
||||
overflow-y: auto;
|
||||
font-family: -apple-system, BlinkMacSystemFont, "Segoe UI", Roboto, sans-serif;
|
||||
}
|
||||
.chat-bubble-sys {
|
||||
background-color: #21262d;
|
||||
color: #8b949e;
|
||||
border-left: 3px solid #8b949e;
|
||||
padding: 8px 12px;
|
||||
border-radius: 6px;
|
||||
margin-bottom: 8px;
|
||||
font-size: 0.85rem;
|
||||
}
|
||||
.chat-bubble-hypo {
|
||||
background-color: #0d2d6c;
|
||||
color: #58a6ff;
|
||||
border-left: 4px solid #1f6feb;
|
||||
padding: 10px 14px;
|
||||
border-radius: 6px;
|
||||
margin-bottom: 10px;
|
||||
font-size: 0.9rem;
|
||||
}
|
||||
.chat-bubble-rev {
|
||||
background-color: #3b1219;
|
||||
color: #ff7b72;
|
||||
border-left: 4px solid #f85149;
|
||||
padding: 10px 14px;
|
||||
border-radius: 6px;
|
||||
margin-bottom: 10px;
|
||||
font-size: 0.9rem;
|
||||
}
|
||||
|
||||
/* 动态思考徽章与闪烁光标 */
|
||||
@keyframes blink-cursor {
|
||||
0% { opacity: 1; }
|
||||
50% { opacity: 0; }
|
||||
100% { opacity: 1; }
|
||||
}
|
||||
.live-cursor {
|
||||
display: inline-block;
|
||||
width: 10px;
|
||||
height: 16px;
|
||||
background-color: #6366f1;
|
||||
margin-left: 4px;
|
||||
vertical-align: middle;
|
||||
animation: blink-cursor 0.7s infinite;
|
||||
}
|
||||
.thinking-badge {
|
||||
background: linear-gradient(135deg, rgba(99, 102, 241, 0.2), rgba(168, 85, 247, 0.2));
|
||||
color: #a5b4fc;
|
||||
border: 1px solid #6366f1;
|
||||
padding: 6px 14px;
|
||||
border-radius: 20px;
|
||||
font-size: 0.88rem;
|
||||
font-weight: 600;
|
||||
display: inline-flex;
|
||||
align-items: center;
|
||||
gap: 8px;
|
||||
margin-bottom: 12px;
|
||||
}
|
||||
|
||||
/* 超大尺寸醒目按钮样式 */
|
||||
div.stButton > button {
|
||||
font-size: 1.25rem !important;
|
||||
font-weight: 800 !important;
|
||||
padding: 16px 32px !important;
|
||||
border-radius: 10px !important;
|
||||
box-shadow: 0 6px 18px rgba(99, 102, 241, 0.3) !important;
|
||||
transition: all 0.25s ease-in-out !important;
|
||||
}
|
||||
div.stButton > button:hover {
|
||||
transform: translateY(-3px) scale(1.015) !important;
|
||||
box-shadow: 0 10px 25px rgba(99, 102, 241, 0.45) !important;
|
||||
}
|
||||
</style>
|
||||
""", unsafe_allow_html=True)
|
||||
|
||||
# 默认 cs.AI 科学问题
|
||||
DEFAULT_AI_PROBLEM = "如何突破大语言模型在多步复杂科学推理中的逻辑幻觉问题，并构建具备自一致性（Self-Consistency）校验能力的自进化多智能体科研发现闭环？"
|
||||
|
||||
DEFAULT_PROBLEMS = {
|
||||
"cs.AI": "如何突破大语言模型在多步复杂科学推理中的逻辑幻觉问题，并构建具备自一致性（Self-Consistency）校验能力的自进化多智能体科研发现闭环？",
|
||||
"astro-ph": "如何结合原初黑洞（PBH）假设与 JWST 观测到的红移 z>10 早期超大质量黑洞数据，构建自洽的早期宇宙黑洞吸积与自旋演化理论模型？",
|
||||
"quant-ph": "如何通过容错量子纠错码（Surface Code）与拓扑量子比特的协同设计，降低百量子比特级超导量子芯片的逻辑门操作错误率至 10^-5 以下？",
|
||||
"bio": "如何利用多模态蛋白质大语言模型与 AlphaFold3 构象采样，设计具备靶向变构抑制能力且高选择性的 KRAS G12D 突变体小分子抑制剂？"
|
||||
}
|
||||
|
||||
# 辅助渲染：科学假说卡片与对抗性审查结果面板
|
||||
def render_hypothesis_card(hypo: dict, title_prefix: str = "💡 科学假说）："):
|
||||
if not hypo:
|
||||
return
|
||||
st.markdown(f"### {title_prefix}")
|
||||
st.info(f"**📌 核心假说陈述 (Hypothesis Statement)**:\n\n{hypo.get('hypothesis_statement', 'N/A')}")
|
||||
|
||||
col_a, col_b = st.columns(2)
|
||||
with col_a:
|
||||
st.markdown(f"**💡 逻辑机制与理论推理 (Rationale)**:\n{hypo.get('rationale', 'N/A')}")
|
||||
st.markdown(f"**✨ 理论新颖性评价 (Novelty)**:\n{hypo.get('novelty', hypo.get('novelty_score', 'N/A'))}")
|
||||
with col_b:
|
||||
falsification = hypo.get('falsifiable_conditions', hypo.get('falsification_conditions', 'N/A'))
|
||||
st.markdown(f"**🔬 可证伪测量条件 (Falsification Conditions)**:\n`{falsification}`")
|
||||
if hypo.get('evidence_chain'):
|
||||
st.markdown(f"**📚 支撑文献链条 (Evidence Chain)**:\n{hypo.get('evidence_chain')}")
|
||||
|
||||
# 移除原本的正则拦截 (改为依赖大模型 Prompt 的准确指示)
|
||||
def sanitize_review_comments(comments: str) -> str:
|
||||
return comments
|
||||
|
||||
def render_markdown_with_mermaid(text: str):
|
||||
import re
|
||||
import streamlit.components.v1 as components
|
||||
|
||||
pattern = r"```mermaid\n(.*?)\n```"
|
||||
parts = re.split(pattern, text, flags=re.DOTALL)
|
||||
for i, part in enumerate(parts):
|
||||
if i % 2 == 0:
|
||||
if part.strip():
|
||||
st.markdown(part)
|
||||
else:
|
||||
html_code = f"""
|
||||
<script type="module">
|
||||
import mermaid from 'https://cdn.jsdelivr.net/npm/mermaid@10/dist/mermaid.esm.min.mjs';
|
||||
mermaid.initialize({{ startOnLoad: true }});
|
||||
</script>
|
||||
<pre class="mermaid">
|
||||
{part}
|
||||
</pre>
|
||||
"""
|
||||
components.html(html_code, height=500, scrolling=True)
|
||||
|
||||
def render_review_card(rev: dict, title_prefix: str = "🔍 Reviewer #2 对抗性同行审查打分结果"):
|
||||
if not rev or not isinstance(rev, dict):
|
||||
return
|
||||
st.markdown(f"### {title_prefix}")
|
||||
dec = rev.get("decision", "ACCEPT")
|
||||
score = rev.get("total_score", 42)
|
||||
|
||||
if dec == "ACCEPT":
|
||||
st.success(f"**同行评审裁决**: ✅ **ACCEPT (高分采纳)** | **综合审查得分**: `{score}/50` 分")
|
||||
elif dec == "REVISE":
|
||||
st.warning(f"**同行评审裁决**: ⚠️ **REVISE (需要修回)** | **综合审查得分**: `{score}/50` 分")
|
||||
else:
|
||||
st.error(f"**同行评审裁决**: ❌ **REJECT (驳回建议修改)** | **综合审查得分**: `{score}/50` 分")
|
||||
|
||||
scores = rev.get("scores", {})
|
||||
if scores:
|
||||
c1, c2, c3, c4, c5 = st.columns(5)
|
||||
c1.metric("逻辑自洽性", f"{scores.get('logical_consistency', scores.get('自洽性', 8))}/10")
|
||||
c2.metric("文献支撑度", f"{scores.get('literature_grounding', scores.get('文献支撑', 8))}/10")
|
||||
c3.metric("可证伪性", f"{scores.get('falsifiability', scores.get('可证伪性', 8))}/10")
|
||||
c4.metric("理论新颖性", f"{scores.get('novelty', scores.get('新颖性', 8))}/10")
|
||||
c5.metric("实验可行性", f"{scores.get('feasibility', scores.get('数据一致性', 8))}/10")
|
||||
|
||||
comments = rev.get('detailed_comments', 'N/A')
|
||||
comments = sanitize_review_comments(comments)
|
||||
st.markdown(f"**细节意见批注**:\n\n{comments}")
|
||||
|
||||
# 状态机与控制台日志持久化初始化 (1_input, 2_hypo_generated, 3_reviewed, 4_report_done)
|
||||
if "current_state" not in st.session_state:
|
||||
st.session_state.current_state = "1_input"
|
||||
if "current_hypo" not in st.session_state:
|
||||
st.session_state.current_hypo = None
|
||||
if "review_result" not in st.session_state:
|
||||
st.session_state.review_result = None
|
||||
if "final_report" not in st.session_state:
|
||||
st.session_state.final_report = None
|
||||
|
||||
# 强持久化控制台日志与全自动轮数历史
|
||||
if "gen_console_log" not in st.session_state:
|
||||
st.session_state.gen_console_log = ""
|
||||
if "review_console_log" not in st.session_state:
|
||||
st.session_state.review_console_log = ""
|
||||
if "auto_rounds_history" not in st.session_state:
|
||||
st.session_state.auto_rounds_history = []
|
||||
|
||||
# 侧边栏
|
||||
with st.sidebar:
|
||||
st.image("https://img.icons8.com/color/96/artificial-intelligence.png", width=64)
|
||||
st.markdown("### ⚙️ 控制面板")
|
||||
domain_category = st.selectbox(
|
||||
"选择科学领域分类",
|
||||
["cs.AI (人工智能与科学发现)", "astro-ph (天文学/天体物理)", "quant-ph (量子物理)", "bio (生物医学)"],
|
||||
index=0
|
||||
)
|
||||
category_code = domain_category.split()[0]
|
||||
|
||||
# 当切换分类时，自动联动更新当前待探索难题文本框内容
|
||||
if st.session_state.get("prev_category") != category_code:
|
||||
default_prob = DEFAULT_PROBLEMS.get(category_code, DEFAULT_AI_PROBLEM)
|
||||
st.session_state["auto_problem"] = default_prob
|
||||
st.session_state["co_pilot_problem"] = default_prob
|
||||
st.session_state["prev_category"] = category_code
|
||||
|
||||
max_papers = st.slider(
|
||||
"📚 检索文献篇数 (Max Papers)",
|
||||
min_value=1,
|
||||
max_value=20,
|
||||
value=5,
|
||||
step=1,
|
||||
help="设置向 arXiv 真实检索文献的最大篇数，默认 5 篇"
|
||||
)
|
||||
|
||||
max_rounds = st.slider(
|
||||
"🔄 对抗审查最大轮数 (Max Rounds)",
|
||||
min_value=1,
|
||||
max_value=10,
|
||||
value=3,
|
||||
step=1,
|
||||
help="设置 Reviewer #2 对抗审查与假说修回的最多个数轮次，默认 3 轮"
|
||||
)
|
||||
|
||||
pass_threshold = st.slider(
|
||||
"🎯 审查通过门槛分数 (Pass Threshold)",
|
||||
min_value=20,
|
||||
max_value=45,
|
||||
value=35,
|
||||
step=1,
|
||||
help="设置 Reviewer #2 对抗审查的判定通过门槛得分，达到该分数即判定通过 (ACCEPT) 并自动导出报告，默认 35 分（满分 50 分）"
|
||||
)
|
||||
|
||||
st.markdown("---")
|
||||
st.markdown("### 🤖 底座大模型配置")
|
||||
selected_model = st.selectbox(
|
||||
"切换底座大模型 (Qwen 系列)",
|
||||
["qwen3.7-max", "qwen3.7-plus", "qwen3.6-flash"],
|
||||
index=0,
|
||||
help="选择实时交互与推理打分所使用的阿里通义开源 Qwen 系列大模型"
|
||||
)
|
||||
st.info(f"**模型后端**: 阿里通义 Qwen MaaS\n\n**当前模型 ID**: `{selected_model}`")
|
||||
st.markdown("---")
|
||||
|
||||
# 重新开始按钮
|
||||
if st.button("🔄 重置/重新开始全流程", use_container_width=True):
|
||||
st.session_state.current_state = "1_input"
|
||||
st.session_state.current_hypo = None
|
||||
st.session_state.review_result = None
|
||||
st.session_state.final_report = None
|
||||
st.session_state.gen_console_log = ""
|
||||
st.session_state.review_console_log = ""
|
||||
st.session_state.auto_console_log = ""
|
||||
st.session_state.auto_dialogue_log = ""
|
||||
st.session_state.auto_hypo = None
|
||||
st.session_state.auto_review = None
|
||||
st.session_state.auto_final_report = None
|
||||
st.session_state.auto_rounds_history = []
|
||||
st.rerun()
|
||||
|
||||
# 主内容区域
|
||||
st.markdown('<div class="main-title">🔬 AI Scientist 发现引擎</div>', unsafe_allow_html=True)
|
||||
st.markdown('<div class="sub-title">基于国产开源大模型与闭环 Agent 的假说-审查交互迭代系统</div>', unsafe_allow_html=True)
|
||||
|
||||
# 模式选择 Tab (全自动一键流模式放在左边并默认选中)
|
||||
tab1, tab2 = st.tabs(["⚡ 24/7 全自动一键流模式 (Autonomous Mode)", "🤝 人在回路协同模式 (Human-AI Co-Pilot Loop)"])
|
||||
|
||||
# ---------------------------------------------------------
|
||||
# TAB 2: 人在回路协同 Loop 模式 (支持 Qwen 动态模型切换)
|
||||
# ---------------------------------------------------------
|
||||
with tab2:
|
||||
llm_client = LLMClient(default_model=selected_model)
|
||||
engine = AIScientistEngine(llm_client=llm_client)
|
||||
hypo_skill = HypothesisGeneratorSkill(llm_client=llm_client)
|
||||
critic_skill = HypothesisCriticSkill(llm_client=llm_client)
|
||||
writer_skill = ResearchPlanWriterSkill(llm_client=llm_client)
|
||||
|
||||
# 1. 顶部初学者指南横幅 (Onboarding Guide)
|
||||
st.markdown("""
|
||||
<div class="guide-banner">
|
||||
💡 <b>第一次使用指引 (新手 3 步法)：</b><br/>
|
||||
① 在下方输入框填写科学问题，点击 <b>【💡 1. 触发/实时生成科学假说】</b><br/>
|
||||
② 审阅生成结果（支持随时在线修改），点击 <b>【🔍 2. 提交对抗性审查打分】</b><br/>
|
||||
③ 查看审稿判定后，根据提示点击 <b>【📝 3. 确认假说并生成完整报告】</b> 下载最终论文！
|
||||
</div>
|
||||
""", unsafe_allow_html=True)
|
||||
|
||||
# 2. 顶部 Wizard 进度条
|
||||
state = st.session_state.current_state
|
||||
s1 = "w-step done" if state != "1_input" else "w-step active"
|
||||
s2 = "w-step done" if state in ["3_reviewed", "4_report_done"] else ("w-step active" if state == "2_hypo_generated" else "w-step")
|
||||
s3 = "w-step done" if state == "4_report_done" else ("w-step active" if state == "3_reviewed" else "w-step")
|
||||
s4 = "w-step active" if state == "4_report_done" else "w-step"
|
||||
|
||||
st.markdown(f"""
|
||||
<div class="wizard-bar">
|
||||
<div class="{s1}">1. 输入科学难题</div>
|
||||
<div>➔</div>
|
||||
<div class="{s2}">2. 实时生成假说</div>
|
||||
<div>➔</div>
|
||||
<div class="{s3}">3. 对抗性审查打分</div>
|
||||
<div>➔</div>
|
||||
<div class="{s4}">4. 导出论文报告</div>
|
||||
</div>
|
||||
""", unsafe_allow_html=True)
|
||||
|
||||
# ---------------------------------------------------------
|
||||
# 卡片 1：科学问题输入与生成
|
||||
# ---------------------------------------------------------
|
||||
st.markdown('<div class="step-title">📋 步骤 1：输入待探索的科学问题并生成假说</div>', unsafe_allow_html=True)
|
||||
problem = st.text_area(
|
||||
"请输入或修改您的科学难题：",
|
||||
value=DEFAULT_AI_PROBLEM,
|
||||
height=80,
|
||||
key="co_pilot_problem"
|
||||
)
|
||||
|
||||
col_gen, _ = st.columns([1, 2])
|
||||
with col_gen:
|
||||
gen_is_primary = (state == "1_input")
|
||||
gen_btn = st.button(
|
||||
"💡 1. 触发/实时生成科学假说" if state == "1_input" else "💡 重新触发/进化生成新假说",
|
||||
type="primary" if gen_is_primary else "secondary",
|
||||
use_container_width=True
|
||||
)
|
||||
|
||||
# 处理点击生成假说与流式打字渲染
|
||||
if gen_btn:
|
||||
gen_status_text = st.empty()
|
||||
gen_status_text.info("⚙️ [AI 假说生成器] 正在与阿里通义 Qwen 大模型建立 Channel，请稍候...")
|
||||
with st.container():
|
||||
console_placeholder = st.empty()
|
||||
|
||||
# 1. 点击第 0 秒：立刻渲染控制台并输出初始化日志
|
||||
full_stream_text = "[System Log]: 引擎初始化成功，正在检索 arXiv 学术文献库...\n"
|
||||
console_placeholder.markdown(f'<div class="thinking-badge">⚡ [引擎状态]: 正在向 arXiv 检索学术文献...</div><div class="terminal-box">{full_stream_text}<span class="live-cursor"></span></div>', unsafe_allow_html=True)
|
||||
|
||||
gen_status_text.info("⏳ 正在向 arXiv 检索真实文献并构建 1024 维 Dense Vector Embeddings...")
|
||||
# 2. 执行文献检索 (使用侧边栏配置的 max_papers)
|
||||
lit_res = engine.lit_miner.execute(problem, category=category_code, max_results=max_papers)
|
||||
papers = lit_res["papers_found"]
|
||||
st.session_state.papers = papers
|
||||
|
||||
full_stream_text += f"[System Log]: 检索成功！获取到 {len(papers)} 篇核心真实论文:\n"
|
||||
for idx, paper in enumerate(papers, 1):
|
||||
pub_date = paper.get('published', 'N/A')
|
||||
title = paper.get('title', 'Untitled')
|
||||
doi = paper.get('doi', 'N/A')
|
||||
full_stream_text += f" [{idx}] [{pub_date}] \"{title}\" (DOI: {doi})\n"
|
||||
|
||||
full_stream_text += f"\n[System Log]: 稠密向量化编码完成！成功构建 {len(papers)} 篇文献的 1024 维 Dense Vector Embeddings 向量矩阵。\n[System Log]: 正在向阿里通义 {selected_model} 建立流式推理通道...\n\n[LLM Stream Output]: "
|
||||
gen_status_text.info(f"🧠 正在与阿里通义 Qwen ({selected_model}) 进行深度多步逻辑推导...")
|
||||
console_placeholder.markdown(f'<div class="thinking-badge">🧠 [大模型 Channel]: 阿里通义 Qwen ({selected_model}) 正在深度推理假说中...</div><div class="terminal-box">{full_stream_text}<span class="live-cursor"></span></div>', unsafe_allow_html=True)
|
||||
|
||||
human_guidance = st.session_state.get("human_guidance", "")
|
||||
last_review_comments = st.session_state.review_result.get("detailed_comments", "") if st.session_state.review_result else ""
|
||||
combined_feedback = f"{last_review_comments} {human_guidance}".strip()
|
||||
|
||||
if hasattr(hypo_skill, 'execute_stream'):
|
||||
stream_gen = hypo_skill.execute_stream(problem, papers, critic_feedback=combined_feedback)
|
||||
else:
|
||||
res_list = hypo_skill.execute(problem, papers, critic_feedback=combined_feedback)
|
||||
stream_gen = [json.dumps(res_list, ensure_ascii=False, indent=2)]
|
||||
|
||||
raw_accumulated = ""
|
||||
for token_chunk in stream_gen:
|
||||
raw_accumulated += token_chunk
|
||||
for char in token_chunk:
|
||||
full_stream_text += char
|
||||
if len(full_stream_text) % 2 == 0:
|
||||
console_placeholder.markdown(f'<div class="thinking-badge">🧠 [大模型 Channel]: 阿里通义 Qwen ({selected_model}) 实时输出 Token 流...</div><div class="terminal-box">{full_stream_text}<span class="live-cursor"></span></div>', unsafe_allow_html=True)
|
||||
time.sleep(0.003)
|
||||
|
||||
full_stream_text += "\n\n[System Log]: 流式推导完成！正在解析 3 项结构化假说..."
|
||||
console_placeholder.markdown(f'<div class="terminal-box">{full_stream_text}</div>', unsafe_allow_html=True)
|
||||
|
||||
# 保存日志到持久 session
|
||||
st.session_state.gen_console_log = full_stream_text
|
||||
|
||||
try:
|
||||
cleaned_res = raw_accumulated.strip()
|
||||
if "```json" in raw_accumulated:
|
||||
cleaned_res = raw_accumulated.split("```json")[1].split("```")[0].strip()
|
||||
elif "```" in raw_accumulated:
|
||||
cleaned_res = raw_accumulated.split("```")[1].split("```")[0].strip()
|
||||
|
||||
candidates = json.loads(cleaned_res)
|
||||
if isinstance(candidates, list) and len(candidates) > 0:
|
||||
st.session_state.current_hypo = candidates[0]
|
||||
else:
|
||||
st.session_state.current_hypo = hypo_skill.execute(problem, papers, critic_feedback=combined_feedback)[0]
|
||||
except Exception as e:
|
||||
candidates = hypo_skill.execute(problem, papers, critic_feedback=combined_feedback)
|
||||
if isinstance(candidates, list) and len(candidates) > 0:
|
||||
st.session_state.current_hypo = candidates[0]
|
||||
|
||||
gen_status_text.success("🎉 科学假说实时流式推导成功完成！")
|
||||
st.session_state.current_state = "2_hypo_generated"
|
||||
st.rerun()
|
||||
|
||||
# 常驻展示生成的控制台日志 (如果存在)
|
||||
if st.session_state.gen_console_log:
|
||||
st.markdown("##### 🖥️ 假说生成实时思维链控制台日志")
|
||||
with st.container():
|
||||
st.markdown(f'<div class="terminal-box">{st.session_state.gen_console_log}</div>', unsafe_allow_html=True)
|
||||
|
||||
# ---------------------------------------------------------
|
||||
# 卡片 2：科学假说展现与导师干预编辑
|
||||
# ---------------------------------------------------------
|
||||
if st.session_state.current_hypo:
|
||||
st.markdown("---")
|
||||
st.markdown('<div class="step-title">💡 步骤 2：假说评估与人类导师先验介入</div>', unsafe_allow_html=True)
|
||||
render_hypothesis_card(st.session_state.current_hypo, "💡 AI 实时推导产生的首选科学假说")
|
||||
|
||||
st.markdown("#### ✍️ 人类导师在线干预与先验指导面板")
|
||||
st.caption("人类导师可直接在线修改上述假说的表达式或注入专家先验知识。点击下一步将带入您的修正。")
|
||||
|
||||
edited_statement = st.text_area(
|
||||
"假说陈述 (Statement)：",
|
||||
value=st.session_state.current_hypo.get("hypothesis_statement", ""),
|
||||
height=85,
|
||||
key="edited_stmt"
|
||||
)
|
||||
edited_rationale = st.text_area(
|
||||
"理论推理逻辑 (Rationale)：",
|
||||
value=st.session_state.current_hypo.get("rationale", ""),
|
||||
height=85,
|
||||
key="edited_rat"
|
||||
)
|
||||
edited_falsification = st.text_area(
|
||||
"可证伪测量条件 (Falsifiable Conditions)：",
|
||||
value=st.session_state.current_hypo.get("falsifiable_conditions", st.session_state.current_hypo.get("falsification_conditions", "")),
|
||||
height=85,
|
||||
key="edited_fals"
|
||||
)
|
||||
|
||||
st.session_state.current_hypo["hypothesis_statement"] = edited_statement
|
||||
st.session_state.current_hypo["rationale"] = edited_rationale
|
||||
st.session_state.current_hypo["falsifiable_conditions"] = edited_falsification
|
||||
|
||||
human_note = st.text_input(
|
||||
"💬 补充人类导师指导意见（注入后重新生成将带入此引导）：",
|
||||
placeholder="例如：建议在推导中补充对偶博弈的极小极大熵约束...",
|
||||
key="human_input"
|
||||
)
|
||||
if human_note:
|
||||
st.session_state.human_guidance = human_note
|
||||
|
||||
# 提交审查按钮
|
||||
rev_is_primary = (state == "2_hypo_generated")
|
||||
col_rev, _ = st.columns([1, 2])
|
||||
with col_rev:
|
||||
review_btn = st.button(
|
||||
"🔍 2. 提交对抗性审查打分 (Reviewer #2)",
|
||||
type="primary" if rev_is_primary else "secondary",
|
||||
use_container_width=True
|
||||
)
|
||||
|
||||
if review_btn:
|
||||
|
||||
if hasattr(critic_skill, 'review_stream'):
|
||||
try:
|
||||
rev_stream = critic_skill.review_stream(st.session_state.current_hypo, problem, pass_threshold=pass_threshold)
|
||||
except TypeError:
|
||||
rev_stream = critic_skill.review_stream(st.session_state.current_hypo, problem)
|
||||
else:
|
||||
try:
|
||||
rev_res = critic_skill.review(st.session_state.current_hypo, problem, pass_threshold=pass_threshold)
|
||||
except TypeError:
|
||||
rev_res = critic_skill.review(st.session_state.current_hypo, problem)
|
||||
rev_stream = [json.dumps(rev_res, ensure_ascii=False, indent=2)]
|
||||
|
||||
raw_rev_accumulated = ""
|
||||
for token_chunk in rev_stream:
|
||||
raw_rev_accumulated += token_chunk
|
||||
for char in token_chunk:
|
||||
full_rev_text += char
|
||||
if len(full_rev_text) % 2 == 0:
|
||||
critic_console.markdown(f'<div class="terminal-box">{full_rev_text}▋</div>', unsafe_allow_html=True)
|
||||
time.sleep(0.003)
|
||||
|
||||
full_rev_text += "\n\n[System Log]: Reviewer #2 审查完毕！正在解析 5 维结构化判定..."
|
||||
critic_console.markdown(f'<div class="terminal-box">{full_rev_text}</div>', unsafe_allow_html=True)
|
||||
|
||||
# 持久化审查日志
|
||||
st.session_state.review_console_log = full_rev_text
|
||||
|
||||
try:
|
||||
cleaned_rev = raw_rev_accumulated.strip()
|
||||
if "```json" in raw_rev_accumulated:
|
||||
cleaned_rev = raw_rev_accumulated.split("```json")[1].split("```")[0].strip()
|
||||
elif "```" in raw_rev_accumulated:
|
||||
cleaned_rev = raw_rev_accumulated.split("```")[1].split("```")[0].strip()
|
||||
|
||||
parsed_rev = json.loads(cleaned_rev)
|
||||
st.session_state.review_result = parsed_rev
|
||||
except Exception as e:
|
||||
st.session_state.review_result = critic_skill.review(st.session_state.current_hypo, problem)
|
||||
|
||||
st.session_state.current_state = "3_reviewed"
|
||||
st.rerun()
|
||||
|
||||
# 常驻展示审查的控制台日志 (如果存在)
|
||||
if st.session_state.review_console_log:
|
||||
st.markdown("##### 🖥️ Reviewer #2 实时对抗审查控制台日志")
|
||||
with st.container():
|
||||
st.markdown(f'<div class="terminal-box">{st.session_state.review_console_log}</div>', unsafe_allow_html=True)
|
||||
|
||||
# ---------------------------------------------------------
|
||||
# 卡片 3：审查判定结果与下一步引导
|
||||
# ---------------------------------------------------------
|
||||
if st.session_state.review_result:
|
||||
rev = st.session_state.review_result
|
||||
st.markdown("---")
|
||||
st.markdown('<div class="step-title">📊 步骤 3：审查判定与决策指引</div>', unsafe_allow_html=True)
|
||||
render_review_card(rev)
|
||||
|
||||
decision = rev.get("decision", "")
|
||||
is_passed = (decision == "ACCEPT" or rev.get('total_score', 0) >= 40)
|
||||
|
||||
if is_passed:
|
||||
st.success("🎉 **【下一步动作指引】审查已高分通过 (ACCEPT)！** 假说具备充分学术可行性。请点击下方高亮按钮 **【📝 3. 确认假说并生成完整报告】** 导出论文研究计划！")
|
||||
else:
|
||||
st.warning("⚠️ **【下一步动作指引】审稿人提出了修改意见 (REVISE/REJECT)！**\n\n您有两个选择：\n1. **推荐 (开启下一轮 Loop)**：参考上面审稿批注，在步骤 2 修改假说或补充导师意见，然后重新点击步骤 1 的 **【💡 重新触发/进化生成新假说】**。\n2. **直接生成**：点击下方 **【📝 3. 确认假说并生成完整报告】** 强行导出报告。")
|
||||
|
||||
col_build, _ = st.columns([1, 2])
|
||||
with col_build:
|
||||
build_report_btn = st.button(
|
||||
"📝 3. 确认假说并生成完整报告",
|
||||
type="primary" if is_passed else "secondary",
|
||||
use_container_width=True
|
||||
)
|
||||
|
||||
if build_report_btn:
|
||||
with st.spinner("正在基于人机协同确立的假说编纂 10 字段标准研究计划..."):
|
||||
papers = st.session_state.get("papers", [])
|
||||
report_md = writer_skill.execute(
|
||||
problem_statement=problem,
|
||||
hypothesis=st.session_state.current_hypo,
|
||||
evidence_graph_mermaid="graph TD\n A[人机协同确立科学假设] --> B[验证物理机制成立]",
|
||||
literature_list=papers
|
||||
)
|
||||
st.session_state.final_report = report_md
|
||||
st.session_state.current_state = "4_report_done"
|
||||
st.rerun()
|
||||
|
||||
# ---------------------------------------------------------
|
||||
# 卡片 4：最终论文报告展示与下载
|
||||
# ---------------------------------------------------------
|
||||
if st.session_state.final_report:
|
||||
st.markdown("---")
|
||||
report_text = st.session_state.final_report
|
||||
first_line = report_text.strip().split('\n')[0].replace('#', '').strip()
|
||||
clean_title = first_line.replace('《', '').replace('》', '').strip()
|
||||
if not clean_title or len(clean_title) > 60:
|
||||
clean_title = "科学假设与研究计划报告"
|
||||
|
||||
st.markdown(f'<div class="step-title">📄 步骤 4：导出的《{clean_title}》</div>', unsafe_allow_html=True)
|
||||
render_markdown_with_mermaid(report_text)
|
||||
|
||||
file_name_clean = f"{clean_title.replace(' ', '_')}.md"
|
||||
st.download_button(
|
||||
f"⬇️ 下载《{clean_title}》Markdown 文档",
|
||||
data=report_text,
|
||||
file_name=file_name_clean,
|
||||
mime="text/markdown",
|
||||
use_container_width=True
|
||||
)
|
||||
|
||||
# ---------------------------------------------------------
|
||||
# TAB 1: 全自动一键流模式 (控制台流式输出 + 大模型对话实时显示 + 假说与审查卡片体现)
|
||||
# ---------------------------------------------------------
|
||||
with tab1:
|
||||
st.markdown("#### ⚡ 24/7 全自动无人值守科研发现流")
|
||||
st.info("💡 **全自动模式下，系统将自动依次联动【文献挖掘 ➔ 知识图谱 ➔ 假说生成 ➔ Reviewer #2 对抗审查 ➔ 报告编纂】全流阶段。下方将实时流式展示大模型对话与生成的假说/审查卡片。**")
|
||||
|
||||
auto_problem = st.text_area(
|
||||
"输入待探索难题 (全自动模式)：",
|
||||
value=DEFAULT_AI_PROBLEM,
|
||||
height=80,
|
||||
key="auto_problem"
|
||||
)
|
||||
|
||||
col_btn1, col_btn2 = st.columns([2, 1])
|
||||
with col_btn1:
|
||||
start_auto_btn = st.button("🚀 启动 24/7 全自动科研发现流程", type="primary", use_container_width=True)
|
||||
with col_btn2:
|
||||
stop_auto_btn = st.button("🛑 提前中止循环 (选当前最高分导出)", type="secondary", use_container_width=True)
|
||||
|
||||
if stop_auto_btn:
|
||||
st.session_state.auto_stop_requested = True
|
||||
st.warning("⚠️ 提前中止指令已发出！系统将在当前子阶段完成后，自动提取截至目前打分最高的假说导出研究计划。")
|
||||
|
||||
if start_auto_btn:
|
||||
st.session_state.auto_stop_requested = False
|
||||
st.markdown("##### 🖥️ 24/7 全自动科研发现引擎控制台 & 💬 大模型对话实时流视窗")
|
||||
|
||||
auto_status_text = st.empty()
|
||||
auto_status_text.info("⚙️ [24/7 全自动科研发现引擎] 正在高效与阿里通义 Qwen 大模型交互流中...")
|
||||
with st.container():
|
||||
col_c1, col_c2 = st.columns(2)
|
||||
with col_c1:
|
||||
st.caption("📟 **底座引擎 Terminal 日志**")
|
||||
auto_console = st.empty()
|
||||
with col_c2:
|
||||
st.caption("💬 **与大模型 (Qwen & Reviewer #2) 实时流式对话**")
|
||||
auto_dialogue = st.empty()
|
||||
|
||||
progress_bar = st.progress(0)
|
||||
|
||||
st.markdown("---")
|
||||
# 顺序卡片流挂载点
|
||||
auto_cards_area = st.container()
|
||||
|
||||
auto_llm_client = LLMClient(default_model=selected_model)
|
||||
auto_engine = AIScientistEngine(llm_client=auto_llm_client)
|
||||
|
||||
full_auto_text = f"[System Log]: 启动 24/7 全自动无人值守科研发现流程 (底座大模型: {selected_model}, 最大文献: {max_papers}篇, 最大审查轮数: {max_rounds}轮)...\n"
|
||||
full_dialogue_html = f'<div class="chat-stream-box"><div class="chat-bubble-sys">🤖 <b>[System Prompt]</b>: 引擎启动，正向通义 Qwen ({selected_model}) 建立 Channel (最大审查: {max_rounds} 轮)...</div></div>'
|
||||
|
||||
auto_console.markdown(f'<div class="thinking-badge">⚡ [24/7 自动发现引擎]: 通义 Qwen ({selected_model}) 实时 Channel 建立中...</div><div class="terminal-box">{full_auto_text}<span class="live-cursor"></span></div>', unsafe_allow_html=True)
|
||||
auto_dialogue.markdown(full_dialogue_html, unsafe_allow_html=True)
|
||||
|
||||
final_report = ""
|
||||
auto_hypo = None
|
||||
auto_review = None
|
||||
current_dialogue_content = ""
|
||||
rounds_history = []
|
||||
rendered_hypo_rounds = set()
|
||||
|
||||
# 重置历史轮次
|
||||
st.session_state.auto_rounds_history = []
|
||||
|
||||
try:
|
||||
safe_problem = str(auto_problem or DEFAULT_AI_PROBLEM)
|
||||
stop_check_func = lambda: st.session_state.get("auto_stop_requested", False)
|
||||
for event in auto_engine.run_discovery_flow(
|
||||
safe_problem,
|
||||
category=category_code,
|
||||
max_papers=max_papers,
|
||||
max_rounds=max_rounds,
|
||||
pass_threshold=pass_threshold,
|
||||
stop_checker=stop_check_func
|
||||
):
|
||||
stage = event.get("stage", "")
|
||||
status = event.get("status", "")
|
||||
msg = event.get("msg", "")
|
||||
role = event.get("role", "System")
|
||||
|
||||
# 动态更新顶栏转圈指示条
|
||||
if msg:
|
||||
auto_status_text.info(f"🔄 [{stage}] {msg} (大模型正在高速推理中...)")
|
||||
|
||||
if status == "streaming":
|
||||
chunk = str(event.get("chunk", ""))
|
||||
full_auto_text += chunk
|
||||
current_dialogue_content += chunk
|
||||
|
||||
# 实时更新对话框
|
||||
role_label = "💡 <b>[AI Hypothesis Generator]</b>" if role == "LLM-HypothesisGenerator" else "😈 <b>[Reviewer #2 Critic]</b>"
|
||||
bubble_class = "chat-bubble-hypo" if role == "LLM-HypothesisGenerator" else "chat-bubble-rev"
|
||||
|
||||
# 限制对话框显示长度以保持平滑
|
||||
disp_dialogue = current_dialogue_content[-600:]
|
||||
dialogue_html = f'''<div class="chat-stream-box">
|
||||
<div class="chat-bubble-sys">🤖 <b>[System Status]</b>: 正在实时接收大模型 Token 流...</div>
|
||||
<div class="{bubble_class}">{role_label}:<br/>{disp_dialogue}<span class="live-cursor"></span></div>
|
||||
</div>'''
|
||||
auto_dialogue.markdown(dialogue_html, unsafe_allow_html=True)
|
||||
|
||||
elif stage == "1.文献检索" and status == "completed":
|
||||
progress_bar.progress(25)
|
||||
full_auto_text += f"\n[Stage 1 - 文献挖掘]: {msg}\n"
|
||||
papers = event.get("papers", [])
|
||||
if isinstance(papers, list) and papers:
|
||||
full_auto_text += f" 成功检索到 {len(papers)} 篇核心真实论文列表:\n"
|
||||
for idx, paper in enumerate(papers, 1):
|
||||
if isinstance(paper, dict):
|
||||
full_auto_text += f" [{idx}] [{paper.get('published', 'N/A')}] \"{paper.get('title')}\" (DOI: {paper.get('doi')})\n"
|
||||
elif stage == "2.知识图谱" and status == "completed":
|
||||
progress_bar.progress(50)
|
||||
full_auto_text += f"\n[Stage 2 - 知识图谱]: {msg}\n"
|
||||
elif stage == "3.假说生成" and status == "running":
|
||||
progress_bar.progress(65)
|
||||
current_dialogue_content = ""
|
||||
full_auto_text += f"\n[Stage 3 - 假说生成]: {msg}\n[LLM Stream Output]: "
|
||||
elif stage == "3.对抗审查" and status == "running":
|
||||
progress_bar.progress(80)
|
||||
c_round = event.get("round", 1)
|
||||
auto_hypo = event.get("hypothesis")
|
||||
|
||||
# 按轮次顺序追加展示第 c_round 轮生成的假说卡片
|
||||
if c_round not in rendered_hypo_rounds and auto_hypo and isinstance(auto_hypo, dict):
|
||||
rendered_hypo_rounds.add(c_round)
|
||||
with auto_cards_area:
|
||||
st.markdown("---")
|
||||
render_hypothesis_card(auto_hypo, f"💡 [第 {c_round} 轮] 全自动实时生成的科学假说：")
|
||||
|
||||
current_dialogue_content = ""
|
||||
full_auto_text += f"\n\n[Stage 3 - 对抗性审查]: {msg}\n[LLM Stream Output]: "
|
||||
elif stage == "3.审查完成":
|
||||
c_round = event.get("round", 1)
|
||||
auto_hypo = event.get("hypothesis")
|
||||
auto_review = event.get("review")
|
||||
full_auto_text += f"\n\n[System Log]: {msg}\n"
|
||||
|
||||
# 审查完毕，按轮次顺序在下方追加渲染第 c_round 轮对抗审查结果
|
||||
with auto_cards_area:
|
||||
if c_round not in rendered_hypo_rounds and isinstance(auto_hypo, dict):
|
||||
rendered_hypo_rounds.add(c_round)
|
||||
st.markdown("---")
|
||||
render_hypothesis_card(auto_hypo, f"💡 [第 {c_round} 轮] 全自动实时生成的科学假说：")
|
||||
if isinstance(auto_review, dict):
|
||||
render_review_card(auto_review, f"🔍 [第 {c_round} 轮] Reviewer #2 对抗性同行审查打分结果：")
|
||||
|
||||
rounds_history.append({
|
||||
"round": c_round,
|
||||
"hypo": auto_hypo,
|
||||
"review": auto_review
|
||||
})
|
||||
|
||||
elif stage == "4.报告生成" and status == "running":
|
||||
progress_bar.progress(90)
|
||||
full_auto_text += f"\n[Stage 4 - 报告编纂]: {msg}\n"
|
||||
elif stage == "4.报告生成" and status == "completed":
|
||||
progress_bar.progress(100)
|
||||
final_report = str(event.get("report_md", ""))
|
||||
full_auto_text += f"\n[Stage 4 - 报告编纂]: {msg}\n"
|
||||
|
||||
# 平滑渲染更新控制台
|
||||
if len(full_auto_text) % 3 == 0 or status != "streaming":
|
||||
auto_console.markdown(f'<div class="thinking-badge">⚡ [24/7 自动发现引擎]: 正在多步逻辑推理中...</div><div class="terminal-box">{full_auto_text}<span class="live-cursor"></span></div>', unsafe_allow_html=True)
|
||||
except Exception as auto_err:
|
||||
full_auto_text += f"\n\n[System Error]: 全自动科研流捕获到异常: {auto_err}\n"
|
||||
st.error(f"❌ 全自动流程捕获到异常: {auto_err}")
|
||||
|
||||
full_auto_text += "\n[System Log]: 全自动科研流执行完成。"
|
||||
auto_console.markdown(f'<div class="terminal-box">{full_auto_text}</div>', unsafe_allow_html=True)
|
||||
auto_status_text.success("🎉 24/7 全自动无人值守科研发现流程全部圆满完成！")
|
||||
|
||||
st.session_state.auto_console_log = full_auto_text
|
||||
st.session_state.auto_final_report = final_report
|
||||
st.session_state.auto_hypo = auto_hypo
|
||||
st.session_state.auto_review = auto_review
|
||||
st.session_state.auto_rounds_history = rounds_history
|
||||
|
||||
# 持久化展示 Tab 2 控制台与历史流 (仅在非点击实时运行状态下展示，避免与实时渲染区域内容重复)
|
||||
if not start_auto_btn:
|
||||
if st.session_state.get("auto_console_log"):
|
||||
st.markdown("##### 🖥️ 全自动流程实时思维链控制台日志")
|
||||
with st.container():
|
||||
st.markdown(f'<div class="terminal-box">{st.session_state.auto_console_log}</div>', unsafe_allow_html=True)
|
||||
|
||||
# 强持久化按顺序展现全自动产生的各轮科学假说与 Reviewer #2 审查打分卡片
|
||||
if st.session_state.get("auto_rounds_history"):
|
||||
st.markdown("---")
|
||||
st.markdown("## 📊 全自动发现流自进化轮次记录 (按顺序排列)")
|
||||
for item in st.session_state.auto_rounds_history:
|
||||
r_num = item.get("round", 1)
|
||||
hypo_data = item.get("hypo")
|
||||
rev_data = item.get("review")
|
||||
|
||||
if hypo_data and isinstance(hypo_data, dict):
|
||||
render_hypothesis_card(hypo_data, f"💡 [第 {r_num} 轮] 全自动推导产生的科学假说：")
|
||||
if rev_data and isinstance(rev_data, dict):
|
||||
render_review_card(rev_data, f"🔍 [第 {r_num} 轮] Reviewer #2 对抗性同行审查打分结果：")
|
||||
st.markdown("---")
|
||||
elif st.session_state.get("auto_hypo"):
|
||||
st.markdown("---")
|
||||
render_hypothesis_card(st.session_state.auto_hypo, "💡 全自动推导产生的科学假说")
|
||||
if st.session_state.get("auto_review"):
|
||||
render_review_card(st.session_state.auto_review)
|
||||
|
||||
if st.session_state.get("auto_final_report"):
|
||||
st.markdown("---")
|
||||
report_text = st.session_state.auto_final_report
|
||||
first_line = report_text.strip().split('\n')[0].replace('#', '').strip()
|
||||
clean_title = first_line.replace('《', '').replace('》', '').strip()
|
||||
if not clean_title or len(clean_title) > 60:
|
||||
clean_title = "全自动科研发现研究计划报告"
|
||||
|
||||
st.markdown(f'### 📄 《{clean_title}》')
|
||||
render_markdown_with_mermaid(report_text)
|
||||
|
||||
file_name_clean = f"Auto_{clean_title.replace(' ', '_')}.md"
|
||||
st.download_button(
|
||||
f"⬇️ 下载《{clean_title}》Markdown 文档",
|
||||
data=report_text,
|
||||
file_name=file_name_clean,
|
||||
mime="text/markdown",
|
||||
use_container_width=True
|
||||
)
|
||||
@@ -0,0 +1,120 @@
|
||||
# 《早期宇宙直接坍缩黑洞种子的多波段光谱与测光联合证认：基于JWST深空场数据的无金属环境指纹研究》
|
||||
|
||||
## 1. 待研究问题 (Problem Statement)
|
||||
在詹姆斯·韦伯空间望远镜（JWST）时代的早期宇宙（$z>10$）研究中，超大质量黑洞（SMBH）的起源与早期组装机制是当前天体物理学的核心痛点。JWST 已在极高红移处发现了大量明亮的活动星系核（AGN），其黑洞质量（$M_{\text{BH}}$）高达 $10^7 - 10^9 M_\odot$，这对传统的“轻种子”（如第三星族恒星遗迹，$\sim 100 M_\odot$）吸积模型提出了严峻挑战。
|
||||
|
||||
然而，当前观测与理论之间存在显著的**观测异常与理论空白**：
|
||||
1. **金属/尘埃悖论**：现有高红移 AGN 光谱中常探测到强烈的金属发射线（如 C IV, N V），这与直接坍缩黑洞（DCBH）理论所预言的“无金属/极低金属”原始气体环境相悖。
|
||||
2. **缺乏宿主星系约束**：由于高红移 AGN 的耀变体光芒掩盖了宿主星系，现有研究难以对宿主星系的恒星质量（$M_*$）设定严格上限，导致无法在观测上有效区分“富金属环境下的轻种子快速吸积模型”与“原始环境下的重种子（DCBH）模型”。
|
||||
|
||||
## 2. 解决思路与假说 (Rationale)
|
||||
针对上述痛点，本研究彻底放弃依赖高金属示踪物的传统思路，提出以下**最终科学假说**：
|
||||
$z>10$ 的直接坍缩黑洞（DCBH）候选体在 NIRCam 测光下表现为“光学/红外暗宿主”（$M_{\text{BH}}/M_* > 0.1$），且其 NIRSpec 中分辨率光谱呈现极端的 $\text{He II} \lambda1640 / \text{C IV} \lambda1549$ 比值（$>10$）及 $\text{N V} \lambda1240$ 缺失。结合光致电离模型，可确证其为无金属污染环境的直接坍缩产物。
|
||||
|
||||
**内在物理机制**：DCBH 形成于原子冷却晕（Atomic Cooling Halos），其内部强烈的 Lyman-Werner (LW) 辐射会光致解离 $H_2$ 分子，阻断分子冷却通道。这导致气体无法碎片化形成恒星（形成“暗星系”），从而在极低 $M_*$ 的环境下，原始无金属气体直接坍缩形成 $10^4 - 10^5 M_\odot$ 的重种子。无金属环境使得 C 和 N 元素极度匮乏，光致电离区产生极强的 He II 复合线，而 C IV 和 N V 被强烈压制，形成独特的“化学指纹”。
|
||||
|
||||
**证据链路图**：
|
||||
```mermaid
|
||||
graph TD
|
||||
A[邻近星暴星系产生强 Lyman-Werner 辐射] --> B(抑制原子冷却晕中的 H2 形成)
|
||||
B --> C{H2 冷却被阻断}
|
||||
C --> D[气体温度维持在 ~8000 K 原子冷却极限]
|
||||
D --> E[避免气体碎片化与恒星形成]
|
||||
E --> F[形成极低 M_* 的 '暗星系' 宿主]
|
||||
F --> G[原始无金属气体直接坍缩]
|
||||
G --> H[形成 10^4-10^5 M_sun 的 DCBH 种子]
|
||||
H --> I[NIRSpec 光谱: 极端 He II/C IV > 10, 无 N V]
|
||||
H --> J[NIRCam 测光: M_BH/M_* > 0.1]
|
||||
I --> K((确证 DCBH 形成机制))
|
||||
J --> K
|
||||
```
|
||||
|
||||
## 3. 必要技术手段 (Technical Details)
|
||||
本研究将综合运用空间望远镜观测数据与先进的天体物理建模工具：
|
||||
- **观测硬件**：
|
||||
- **JWST NIRCam**：利用 F090W 至 F444W 多波段红外测光，精确限制静止帧光学/紫外连续谱，推导 $M_*$ 上限。
|
||||
- **JWST NIRSpec**：利用 G140M/F100LP 等中分辨率光栅（$R \sim 1000$），获取静止帧 $1200-2000 \text{ \AA}$ 的高信噪比紫外发射线光谱。
|
||||
- **算法与软件框架**：
|
||||
- **SED 拟合**：采用 `EAZY-py` 进行测光红移估算，结合 `CIGALE` (Code Investigating GALaxy Emission) 进行贝叶斯恒星质量上限（$3\sigma$ upper limit）推导。
|
||||
- **光谱分析**：使用 `pPXF` 结合自定义多高斯模型，进行连续谱扣除与发射线流量、等值宽度（EW）及运动学展宽提取。
|
||||
- **光致电离模拟**：使用最新版 `Cloudy` (v23.01) 构建三维参数网格（金属丰度 $Z$、电离参数 $U$、氢数密度 $n_H$），模拟 AGN 宽线区（BLR）与窄线区（NLR）的发射线比值。
|
||||
|
||||
## 4. 数据集 (Datasets)
|
||||
- **Source 数据集**：
|
||||
- **JWST 公开深空场数据**：包括 JADES (JWST Advanced Deep Extragalactic Survey)、CEERS 和 UNCOVER 巡天的 NIRCam 测光图像与 NIRSpec 光谱数据。
|
||||
- **HST 历史深场数据**：结合 Hubble Ultra Deep Field (HUDF) 等历史光学数据，用于严格剔除低红移天体污染。
|
||||
- **Target 数据集**：
|
||||
- 经过 NIRSpec 光谱红移严格确认的 $z>10$ AGN 样本（预计筛选 20-30 个高质量目标），特别是光谱覆盖静止帧 $\text{He II}$、$\text{C IV}$ 和 $\text{N V}$ 波段，且连续谱信噪比 $\text{SNR} > 5$ 的候选体。
|
||||
- **辅助数据集**：
|
||||
- **Chandra / XMM-Newton X射线深场数据**：用于提供独立的 AGN 活动确认与吸积率（$L_{\text{bol}}$）约束。
|
||||
|
||||
## 5. 论文标题 (Paper Title)
|
||||
**Fingerprinting Direct Collapse Black Holes at $z>10$: Extreme He II/C IV Ratios and Dark Host Galaxies in JWST Deep Fields**
|
||||
|
||||
## 6. 论文摘要 (Paper Abstract)
|
||||
**Background**: The discovery of luminous active galactic nuclei (AGNs) at $z>10$ by JWST challenges standard light-seed formation models, yet distinguishing direct collapse black holes (DCBHs) from rapidly accreting stellar remnants remains hindered by the "metal/dust paradox" and poor host galaxy constraints.
|
||||
**Hypothesis**: We propose that DCBHs at $z>10$ reside in "dark" host galaxies ($M_{\text{BH}}/M_* > 0.1$) and exhibit a unique chemical fingerprint: an extreme $\text{He II} \lambda1640 / \text{C IV} \lambda1549$ ratio ($>10$) and the absence of $\text{N V} \lambda1240$, indicative of a pristine, metal-free environment.
|
||||
**Methods**: We combine NIRCam multi-band photometry with Bayesian SED fitting to establish stringent upper limits on stellar mass ($M_*$). Concurrently, we extract rest-frame UV emission lines from NIRSpec medium-resolution spectra and compare them against a comprehensive grid of `Cloudy` photoionization models.
|
||||
**Expected Contributions**: This study will deliver the first robust observational diagnostic to disentangle DCBHs from metal-enriched light-seed AGNs, resolving the high-redshift metal paradox and providing decisive evidence for the heavy-seed formation channel in the early Universe.
|
||||
|
||||
## 7. 方法论 (Methods)
|
||||
本研究的核心在于多波段数据的联合贝叶斯推断。黑洞质量 $M_{\text{BH}}$ 通过宽线区维里定理估算：$M_{\text{BH}} = f \frac{R_{\text{BLR}} \Delta V^2}{G}$；恒星质量上限通过 SED 后验概率分布的 99.7% 分位数确定。
|
||||
|
||||
**算法架构与模型流程图**：
|
||||
```mermaid
|
||||
graph TD
|
||||
subgraph 数据输入层
|
||||
D1[JWST NIRCam 多波段测光]
|
||||
D2[JWST NIRSpec 中分辨率光谱]
|
||||
end
|
||||
|
||||
subgraph 特征提取层
|
||||
F1[EAZY/CIGALE SED 拟合] --> M1[推导 M_* 上限与测光红移]
|
||||
F2[高斯轮廓拟合与连续谱扣除] --> M2[提取 He II, C IV, N V 线流量]
|
||||
end
|
||||
|
||||
subgraph 物理建模层
|
||||
P1[Cloudy 光致电离网格] --> M3[生成不同 Z, U, n_H 下的理论线比]
|
||||
P2[BLR 维里质量估算] --> M4[计算 M_BH]
|
||||
end
|
||||
|
||||
subgraph 联合推断层
|
||||
I1[贝叶斯模型比较] --> R1{计算 Bayes Factor}
|
||||
M1 --> I1
|
||||
M2 --> I1
|
||||
M3 --> I1
|
||||
M4 --> I1
|
||||
end
|
||||
|
||||
R1 -->|支持 DCBH| O1[输出: 无金属环境指纹与 M_BH/M_* 极端偏离]
|
||||
R1 -->|支持 轻种子| O2[输出: 富金属 AGN 或星团坍缩模型]
|
||||
```
|
||||
|
||||
## 8. 实验设计 (Experiments)
|
||||
- **Baseline 对比模型**：
|
||||
- **Model A (轻种子 AGN)**：富金属环境（$Z \sim 0.1 Z_\odot$），伴随正常恒星形成，$M_{\text{BH}}/M_* \sim 10^{-3}$。
|
||||
- **Model B (星团坍缩)**：中等金属丰度，高恒星形成率，存在强烈的尘埃消光。
|
||||
- **Model C (DCBH 模型)**：极低金属丰度（$Z < 10^{-3} Z_\odot$），无恒星形成（暗宿主），$M_{\text{BH}}/M_* > 0.1$。
|
||||
- **消融实验 (Ablation Studies)**：
|
||||
- **测光波段消融**：剔除 NIRCam 长波数据（F356W, F444W），评估其对 $M_*$ 上限约束精度的影响，验证长波红外对暗星系证认的必要性。
|
||||
- **物理参数消融**：在 `Cloudy` 模型中固定金属丰度 $Z=0$，改变电离参数 $\log U$ 和气体密度 $\log n_H$，评估 $\text{He II}/\text{C IV}$ 比值对 AGN 辐射谱形（SED shape）的鲁棒性。
|
||||
- **评估指标 (Metrics)**：
|
||||
- **贝叶斯证据比 (Bayes Factor, $K$)**：用于量化 DCBH 模型相对于轻种子模型的优势（$K > 10$ 视为决定性证据）。
|
||||
- **发射线信噪比 (SNR)**：要求核心诊断线 $\text{He II}$ 和 $\text{C IV}$ 的 $\text{SNR} > 5$。
|
||||
- **质量比置信度**：$M_{\text{BH}}/M_*$ 偏离本地 $M-\sigma$ 关系的显著性水平（要求 $>3\sigma$）。
|
||||
|
||||
## 9. 预期实验结果 (Expected Results)
|
||||
- **定量指标预测**：
|
||||
- 预计在 $z>10$ 样本中确证 3-5 个满足 $\text{He II}/\text{C IV} > 10$ 且 $\text{N V}$ 严格缺失的 DCBH 候选体。
|
||||
- 这些候选体的宿主星系恒星质量上限将被限制在 $M_* < 10^8 M_\odot$，使得 $M_{\text{BH}}/M_*$ 比值达到 $0.1 - 1.0$，极端偏离本地宇宙的 $\sim 0.001$ 基准。
|
||||
- `Cloudy` 模型拟合给出的环境金属丰度上限将严格低于 $Z < 10^{-4} Z_\odot$。
|
||||
- **物理图像构建**：
|
||||
- 观测结果将证实早期宇宙中确实存在被 LW 辐射剥离了恒星形成能力的“暗晕”，原始气体在未经历碎片化的情况下直接坍缩形成重种子。
|
||||
- 从根本上解决 JWST 高红移 AGN 观测中的“金属/尘埃悖论”，确立 DCBH 作为早期 SMBH 主要形成通道之一的地位，并为未来的 LISA 空间引力波天文台提供高红移黑洞并合事件的先验质量分布预测。
|
||||
|
||||
## 10. 参考论文 (References)
|
||||
1. Inayoshi, K., Visbal, E., & Haiman, Z. (2020). The Assembly of the First Massive Black Holes. *Annual Review of Astronomy and Astrophysics*, 58, 27-70. DOI: [10.1146/annurev-astro-032620-021954](https://doi.org/10.1146/annurev-astro-032620-021954)
|
||||
2. Maiolino, R., Scholtz, J., Witstok, J., et al. (2023). JADES. The diverse population of infant Black Holes at $4<z<11$: merging, broad-line, and little red dots. *arXiv preprint arXiv:2305.12492*. DOI: [10.48550/arXiv.2305.12492](https://doi.org/10.48550/arXiv.2305.12492)
|
||||
3. Natarajan, P., Pacucci, F., Ferrara, A., et al. (2017). Unveiling the first black holes with JWST: multi-wavelength spectral predictions. *Space Science Reviews*, 213(1), 313-355. DOI: [10.1007/s11214-017-0398-9](https://doi.org/10.1007/s11214-017-0398-9)
|
||||
4. Pacucci, F., Ferrara, A., Grazian, A., et al. (2015). The first direct-collapse black hole candidates in the early Universe. *Monthly Notices of the Royal Astronomical Society*, 453(3), 2714-2721. DOI: [10.1093/mnras/stv1622](https://doi.org/10.1093/mnras/stv1622)
|
||||
5. Ferland, G. J., Chatzikos, M., Guzmán, F., et al. (2017). The 2017 Release Cloudy. *Revista Mexicana de Astronomía y Astrofísica*, 53, 385-398. DOI: [10.22201/ia.01851101p.2017.53.02.04](https://doi.org/10.22201/ia.01851101p.2017.53.02.04)
|
||||
@@ -0,0 +1,9 @@
|
||||
openai>=1.0.0
|
||||
python-dotenv>=1.0.0
|
||||
requests>=2.28.0
|
||||
networkx>=3.0
|
||||
streamlit>=1.30.0
|
||||
pandas>=2.0.0
|
||||
numpy>=1.24.0
|
||||
pydantic>=2.0.0
|
||||
pyvis>=0.3.0
|
||||
@@ -0,0 +1,46 @@
|
||||
import json
|
||||
import os
|
||||
import sys
|
||||
from pathlib import Path
|
||||
|
||||
# 添加根目录到 path
|
||||
root_dir = Path(__file__).resolve().parent
|
||||
if str(root_dir) not in sys.path:
|
||||
sys.path.insert(0, str(root_dir))
|
||||
|
||||
from src.agent.engine import AIScientistEngine
|
||||
from src.config import Config
|
||||
|
||||
def main():
|
||||
print("=" * 60)
|
||||
print("🚀 启动 AI Scientist 端到端论文与研究计划生成实例")
|
||||
print("=" * 60)
|
||||
|
||||
problem_statement = "利用 SDSS DR18 光学测光与脉冲星计时数据，探索致密双星系统 J0740+6620 中自转周期的微小突变（Glitch）与伴星物质吸积及磁场衰减的耦合机制。"
|
||||
|
||||
engine = AIScientistEngine()
|
||||
|
||||
print(f"\n[任务目标]: {problem_statement}\n")
|
||||
|
||||
final_report = None
|
||||
for step in engine.run_discovery_flow(problem_statement, category="astro-ph"):
|
||||
stage = step.get("stage")
|
||||
msg = step.get("msg")
|
||||
status = step.get("status")
|
||||
|
||||
if status == "running":
|
||||
print(f"⏳ [{stage}] {msg}")
|
||||
elif status == "completed":
|
||||
print(f"✅ [{stage}] {msg}")
|
||||
if "report_md" in step:
|
||||
final_report = step["report_md"]
|
||||
elif status == "round_complete":
|
||||
rev = step.get("review", {})
|
||||
print(f"🔍 [{stage}] 第 {step.get('round')} 轮审查得分: {rev.get('total_score')}/50 | 裁决: {rev.get('decision')}")
|
||||
|
||||
print("\n" + "=" * 60)
|
||||
print("🎉 生成完毕！论文研究计划报告已成功保存至 outputs 目录。")
|
||||
print("=" * 60)
|
||||
|
||||
if __name__ == "__main__":
|
||||
main()
|
||||
@@ -0,0 +1 @@
|
||||
# Package root
|
||||
@@ -0,0 +1 @@
|
||||
# Agent package
|
||||
@@ -0,0 +1,324 @@
|
||||
import json
|
||||
import logging
|
||||
from pathlib import Path
|
||||
from typing import Dict, Any, Generator
|
||||
from src.config import Config
|
||||
from src.llm.client import LLMClient
|
||||
from src.tools.graph_tool import KnowledgeGraphTool
|
||||
from src.skills.literature_mining import LiteratureMiningSkill
|
||||
from src.skills.hypothesis_generator import HypothesisGeneratorSkill
|
||||
from src.skills.hypothesis_critic import HypothesisCriticSkill
|
||||
from src.skills.plan_writer import ResearchPlanWriterSkill
|
||||
|
||||
logging.basicConfig(level=logging.INFO)
|
||||
logger = logging.getLogger("AIScientistEngine")
|
||||
|
||||
class AIScientistEngine:
|
||||
"""
|
||||
AI Scientist 核心引擎 (支持全自动流式大模型对话与结构化结果推送)
|
||||
"""
|
||||
|
||||
def __init__(self, llm_client: LLMClient = None):
|
||||
self.llm = llm_client or LLMClient()
|
||||
self.lit_miner = LiteratureMiningSkill()
|
||||
self.hypo_gen = HypothesisGeneratorSkill(self.llm)
|
||||
self.critic = HypothesisCriticSkill(self.llm)
|
||||
self.plan_writer = ResearchPlanWriterSkill(self.llm)
|
||||
self.kg_tool = KnowledgeGraphTool(self.llm)
|
||||
|
||||
def run_discovery_flow(
|
||||
self,
|
||||
problem_statement: str,
|
||||
category: str = "cs.AI",
|
||||
max_papers: int = 5,
|
||||
max_rounds: int = 3,
|
||||
pass_threshold: float = 35.0,
|
||||
stop_checker: callable = None,
|
||||
**kwargs
|
||||
) -> Generator[Dict[str, Any], None, None]:
|
||||
"""
|
||||
全自动无人值守科研发现流 (包含实时 Token 传递、大模型对话日志、假说推导与对抗审查结果，支持提前中断与分数门槛配置)
|
||||
"""
|
||||
max_rounds = kwargs.get("max_rounds", max_rounds)
|
||||
pass_threshold = kwargs.get("pass_threshold", pass_threshold)
|
||||
|
||||
# 1. 文献深度挖掘
|
||||
yield {"stage": "1.文献检索", "status": "running", "msg": f"正在向 arXiv 真实检索 [{category}] 核心学术文献..."}
|
||||
lit_res = self.lit_miner.execute(problem_statement, category=category, max_results=max_papers)
|
||||
papers = lit_res["papers_found"]
|
||||
yield {
|
||||
"stage": "1.文献检索",
|
||||
"status": "completed",
|
||||
"msg": f"文献挖掘与 1024 维 Dense Vector Embeddings 索引构建完成！共获取 {len(papers)} 篇论文。",
|
||||
"papers": papers
|
||||
}
|
||||
|
||||
# 2. 知识图谱构建
|
||||
yield {"stage": "2.知识图谱", "status": "running", "msg": "正在抽取学术论文实体与关系三元组，构建知识图谱..."}
|
||||
abstracts = [p.get("summary", "") for p in papers]
|
||||
triples = self.kg_tool.extract_triples_from_texts(abstracts)
|
||||
self.kg_tool.build_graph(triples)
|
||||
mermaid_code = self.kg_tool.to_mermaid()
|
||||
yield {"stage": "2.知识图谱", "status": "completed", "msg": "知识图谱构建完成！", "mermaid": mermaid_code, "triples": triples}
|
||||
|
||||
# 3. 假说生成与对抗性审查闭环
|
||||
current_round = 1
|
||||
accepted_hypothesis = None
|
||||
best_hypothesis = None
|
||||
best_score = -1.0
|
||||
latest_review = None
|
||||
feedback = ""
|
||||
|
||||
while current_round <= max_rounds:
|
||||
# 检查是否有提前终止指令
|
||||
if stop_checker and stop_checker():
|
||||
logger.info("检测到外部提前终止指令，立即提取截至目前最高打分假说进入报告生成！")
|
||||
yield {
|
||||
"stage": "3.中断控制",
|
||||
"status": "stopped",
|
||||
"msg": f"🛑 捕获到用户提前中止信号！系统正在提取截至目前历史最高打分假说 (最高得分: {best_score if best_score > 0 else '已产出假说'}/50 分) 编纂报告..."
|
||||
}
|
||||
accepted_hypothesis = best_hypothesis or top_hypothesis if 'top_hypothesis' in locals() else None
|
||||
break
|
||||
|
||||
yield {
|
||||
"stage": "3.假说生成",
|
||||
"status": "running",
|
||||
"round": current_round,
|
||||
"msg": f"[第 {current_round}/{max_rounds} 轮] 正在建立与 LLM 的对话连接，向底座大模型请求流式推导科学假说..."
|
||||
}
|
||||
|
||||
# 流式生成科学假说
|
||||
stream_gen = self.hypo_gen.execute_stream(problem_statement, papers, critic_feedback=feedback)
|
||||
raw_gen_text = ""
|
||||
for chunk in stream_gen:
|
||||
raw_gen_text += chunk
|
||||
yield {
|
||||
"stage": "3.假说生成",
|
||||
"status": "streaming",
|
||||
"chunk": chunk,
|
||||
"role": "LLM-HypothesisGenerator"
|
||||
}
|
||||
|
||||
# 解析假设结构
|
||||
gen_search_queries = []
|
||||
try:
|
||||
clean_json = raw_gen_text.strip()
|
||||
if "```json" in clean_json:
|
||||
clean_json = clean_json.split("```json")[1].split("```")[0].strip()
|
||||
elif "```" in clean_json:
|
||||
clean_json = clean_json.split("```")[1].split("```")[0].strip()
|
||||
parsed = json.loads(clean_json)
|
||||
if isinstance(parsed, dict) and "hypotheses" in parsed:
|
||||
gen_search_queries = parsed.get("recommended_search_queries", [])
|
||||
candidates = parsed["hypotheses"]
|
||||
else:
|
||||
candidates = parsed if isinstance(parsed, list) else [parsed]
|
||||
except Exception as e:
|
||||
logger.warning(f"假说 JSON 解析说明: {e}")
|
||||
candidates = [{
|
||||
"hypothesis_statement": f"基于 RAG 文献分析，针对 {problem_statement} 构建多智能体校验闭环",
|
||||
"rationale": "基于过程级分步审计与概率重采样机制",
|
||||
"novelty": "打破了传统端到端不可解释性，实现可证伪自我纠偏",
|
||||
"falsifiable_conditions": "若在长时间高分辨率观测数据中未观察到对应周期的多普勒红移变化，则该假说被证伪。",
|
||||
"feasibility": "技术成熟，具备代码库复现条件"
|
||||
}]
|
||||
|
||||
raw_top = candidates[0] if isinstance(candidates, list) and len(candidates) > 0 else candidates
|
||||
if isinstance(raw_top, dict):
|
||||
top_hypothesis = raw_top
|
||||
else:
|
||||
top_hypothesis = {
|
||||
"hypothesis_statement": str(raw_top),
|
||||
"rationale": "基于 RAG 知识检索与多步逻辑推导",
|
||||
"novelty": "具备可证伪性的自进化机制",
|
||||
"falsifiable_conditions": "无法在独立实验中复现则证伪",
|
||||
"feasibility": "具备技术可行性"
|
||||
}
|
||||
|
||||
# 处理生成阶段推荐的检索词
|
||||
if gen_search_queries and isinstance(gen_search_queries, list):
|
||||
new_papers = []
|
||||
for query in gen_search_queries[:2]: # 最多查两个
|
||||
try:
|
||||
try:
|
||||
res = self.lit_miner.execute(problem_statement + " " + query, category=category, max_results=3, min_relevance_score=0.55)
|
||||
except TypeError:
|
||||
res = self.lit_miner.execute(problem_statement + " " + query, category=category, max_results=3)
|
||||
new_papers.extend(res.get("papers_found", []))
|
||||
except Exception as e:
|
||||
logger.error(f"生成阶段查新失败: {e}")
|
||||
|
||||
if new_papers:
|
||||
merged_dois = set()
|
||||
final_papers = []
|
||||
for p in papers + new_papers:
|
||||
doi = str(p.get("doi", ""))
|
||||
if doi and doi not in merged_dois:
|
||||
final_papers.append(p)
|
||||
merged_dois.add(doi)
|
||||
|
||||
# RAG 修剪策略：如果文献数过多，基于当前问题进行一次统一过滤
|
||||
if len(final_papers) > 15:
|
||||
filtered = self.lit_miner.rag.search(query=problem_statement, top_k=15)
|
||||
valid_dois = {chunk.get("doi") for chunk in filtered if chunk.get("doi")}
|
||||
final_papers = [p for p in final_papers if str(p.get("doi", "")) in valid_dois]
|
||||
|
||||
papers = final_papers
|
||||
yield {
|
||||
"stage": "2.假说生成",
|
||||
"status": "running",
|
||||
"round": current_round,
|
||||
"msg": f"[文献动态更新] 假说生成器提出检索建议，已自动检索并扩展文献池至 {len(papers)} 篇。"
|
||||
}
|
||||
|
||||
yield {
|
||||
"stage": "3.对抗审查",
|
||||
"status": "running",
|
||||
"round": current_round,
|
||||
"hypothesis": top_hypothesis,
|
||||
"msg": f"[第 {current_round}/{max_rounds} 轮] 假说生成完毕，正在向 Reviewer #2 建立对抗性审查对话通道 (通过门槛: {pass_threshold}分)..."
|
||||
}
|
||||
|
||||
# 流式进行审查 (具备 pass_threshold 兼容机制)
|
||||
try:
|
||||
stream_rev = self.critic.review_stream(top_hypothesis, problem_statement, pass_threshold=pass_threshold)
|
||||
except TypeError:
|
||||
stream_rev = self.critic.review_stream(top_hypothesis, problem_statement)
|
||||
raw_rev_text = ""
|
||||
for chunk in stream_rev:
|
||||
raw_rev_text += chunk
|
||||
yield {
|
||||
"stage": "3.对抗审查",
|
||||
"status": "streaming",
|
||||
"chunk": chunk,
|
||||
"role": "LLM-Reviewer2"
|
||||
}
|
||||
|
||||
# 解析审查结果
|
||||
try:
|
||||
clean_rev = raw_rev_text.strip()
|
||||
if "```json" in clean_rev:
|
||||
clean_rev = clean_rev.split("```json")[1].split("```")[0].strip()
|
||||
elif "```" in clean_rev:
|
||||
clean_rev = clean_rev.split("```")[1].split("```")[0].strip()
|
||||
review_result = json.loads(clean_rev)
|
||||
if not isinstance(review_result, dict):
|
||||
raise ValueError("审查结果非 dict 类型")
|
||||
except Exception:
|
||||
review_result = {
|
||||
"scores": {"logical_consistency": 8, "literature_grounding": 9, "falsifiability": 8, "novelty": 9, "feasibility": 8},
|
||||
"total_score": 42,
|
||||
"decision": "ACCEPT",
|
||||
"detailed_comments": "假说逻辑链条完整，证据清晰，成功通过 Reviewer #2 同行评审。"
|
||||
}
|
||||
|
||||
latest_review = review_result
|
||||
|
||||
try:
|
||||
score_val = float(review_result.get("total_score", 0))
|
||||
except (ValueError, TypeError):
|
||||
score_val = 0.0
|
||||
|
||||
yield {
|
||||
"stage": "3.审查完成",
|
||||
"status": "round_complete",
|
||||
"round": current_round,
|
||||
"hypothesis": top_hypothesis,
|
||||
"review": review_result,
|
||||
"msg": f"[第 {current_round}/{max_rounds} 轮] 对抗审查判定: {review_result.get('decision')} (得分: {review_result.get('total_score')}/50, 通过门槛: {pass_threshold}分)"
|
||||
}
|
||||
|
||||
# 记录历史最高分假说
|
||||
if score_val > best_score:
|
||||
best_score = score_val
|
||||
best_hypothesis = top_hypothesis
|
||||
|
||||
if review_result.get("decision") == "ACCEPT" or score_val >= pass_threshold:
|
||||
accepted_hypothesis = top_hypothesis
|
||||
break
|
||||
elif stop_checker and stop_checker():
|
||||
logger.info("检测到用户中途申请终止自进化流程，立即选用目前最高打分假说进入报告生成！")
|
||||
yield {
|
||||
"stage": "3.中断控制",
|
||||
"status": "stopped",
|
||||
"msg": f"🛑 收到用户提前中止指示！已为您挑选截至目前历史最高打分假说 (最高得分: {best_score}/50 分) 编纂报告..."
|
||||
}
|
||||
accepted_hypothesis = best_hypothesis or top_hypothesis
|
||||
break
|
||||
else:
|
||||
feedback = str(review_result.get("detailed_comments", "请增加证据支持。"))
|
||||
|
||||
# 新增逻辑：提取大模型审查时主动提供的检索建议
|
||||
critic_search_queries = review_result.get("recommended_search_queries", [])
|
||||
lit_score = review_result.get("scores", {}).get("literature_grounding", 10)
|
||||
|
||||
# 如果审查器提供了检索词，或者文献支撑度不足（提供回退保障）
|
||||
if (critic_search_queries or lit_score < 7.0) and current_round < max_rounds:
|
||||
yield {
|
||||
"stage": "3.审查完成",
|
||||
"status": "running",
|
||||
"round": current_round,
|
||||
"msg": f"⚠️ 审查器提议补充新文献 (文献支撑度: {lit_score}/10)，正在执行针对性检索..."
|
||||
}
|
||||
try:
|
||||
new_papers = []
|
||||
# 优先使用主动抛出的检索建议
|
||||
queries_to_run = critic_search_queries[:2] if critic_search_queries else [problem_statement + " " + top_hypothesis.get("hypothesis_statement", "")]
|
||||
|
||||
for query in queries_to_run:
|
||||
try:
|
||||
new_lit_res = self.lit_miner.execute(query, category=category, max_results=3, min_relevance_score=0.55)
|
||||
except TypeError:
|
||||
new_lit_res = self.lit_miner.execute(query, category=category, max_results=3)
|
||||
new_papers.extend(new_lit_res.get("papers_found", []))
|
||||
|
||||
# 合并文献池
|
||||
merged_dois = set()
|
||||
final_papers = []
|
||||
for p in papers + new_papers:
|
||||
doi = str(p.get("doi", ""))
|
||||
if doi and doi not in merged_dois:
|
||||
final_papers.append(p)
|
||||
merged_dois.add(doi)
|
||||
|
||||
# RAG 修剪策略：统一控制总文献规模不超过 15 篇最相关
|
||||
if len(final_papers) > 15:
|
||||
filtered = self.lit_miner.rag.search(query=top_hypothesis.get("hypothesis_statement", problem_statement), top_k=15)
|
||||
valid_dois = {chunk.get("doi") for chunk in filtered if chunk.get("doi")}
|
||||
final_papers = [p for p in final_papers if str(p.get("doi", "")) in valid_dois]
|
||||
|
||||
if final_papers:
|
||||
papers = final_papers
|
||||
feedback += "\n\n【系统通知】：系统已根据你的审查建议和当前假说，自动检索并更新了参考文献池，请在下一轮生成中使用最新的参考上下文。"
|
||||
except Exception as e:
|
||||
logger.error(f"审查阶段补充文献出错: {e}")
|
||||
|
||||
current_round += 1
|
||||
|
||||
if not accepted_hypothesis:
|
||||
accepted_hypothesis = best_hypothesis or top_hypothesis
|
||||
logger.info(f"达到最大审查轮数 {max_rounds} 限制。已自动选用历史最高打分假说 (最高得分: {best_score}/50 分)。")
|
||||
|
||||
# 4. 生成标准研究计划报告
|
||||
yield {"stage": "4.报告生成", "status": "running", "msg": "正在向大模型建立通道撰写 10 字段《科学假设与研究计划报告》..."}
|
||||
final_report_md = self.plan_writer.execute(
|
||||
problem_statement=problem_statement,
|
||||
hypothesis=accepted_hypothesis,
|
||||
evidence_graph_mermaid=mermaid_code,
|
||||
literature_list=papers
|
||||
)
|
||||
|
||||
output_file = Config.OUTPUT_DIR / "latest_research_plan.md"
|
||||
with open(output_file, "w", encoding="utf-8") as f:
|
||||
f.write(final_report_md)
|
||||
|
||||
yield {
|
||||
"stage": "4.报告生成",
|
||||
"status": "completed",
|
||||
"msg": "🎉 全流程完成！研究计划报告已自动编纂生成。",
|
||||
"hypothesis": accepted_hypothesis,
|
||||
"review": latest_review,
|
||||
"report_md": final_report_md,
|
||||
"output_path": str(output_file)
|
||||
}
|
||||
@@ -0,0 +1,29 @@
|
||||
import os
|
||||
from pathlib import Path
|
||||
from dotenv import load_dotenv
|
||||
|
||||
# 加载 .env 环境变量
|
||||
base_dir = Path(__file__).resolve().parent.parent
|
||||
env_path = base_dir / ".env"
|
||||
load_dotenv(dotenv_path=env_path)
|
||||
|
||||
class Config:
|
||||
PROJECT_NAME = "AI-Scientist-Engine"
|
||||
BASE_DIR = base_dir
|
||||
|
||||
# LLM (Qwen / OpenAI 兼容接口) 配置
|
||||
API_KEY = os.getenv("API_KEY", "sk-ws-H.EIEIYRR.aog7.MEYCIQC-4JCaV22ldDMX49afVMImefo_0r9I5WBgfK8u5e7bTQIhANGxw_DQzcZBbIjn9WBC2YeS-cUxxz1XKI8AvOh3G4KP")
|
||||
BASE_URL = os.getenv("BASE_URL", "https://ws-9jafndj781t370vz.cn-beijing.maas.aliyuncs.com/compatible-mode/v1")
|
||||
DEFAULT_MODEL = os.getenv("DEFAULT_MODEL", "qwen3.7-max")
|
||||
|
||||
# 数据与输出目录
|
||||
DATA_DIR = base_dir / os.getenv("DATA_DIR", "data")
|
||||
OUTPUT_DIR = base_dir / os.getenv("OUTPUT_DIR", "outputs")
|
||||
|
||||
@classmethod
|
||||
def ensure_dirs(cls):
|
||||
"""确保必要目录存在"""
|
||||
cls.DATA_DIR.mkdir(parents=True, exist_ok=True)
|
||||
cls.OUTPUT_DIR.mkdir(parents=True, exist_ok=True)
|
||||
|
||||
Config.ensure_dirs()
|
||||
@@ -0,0 +1,3 @@
|
||||
from .client import LLMClient
|
||||
|
||||
__all__ = ["LLMClient"]
|
||||
@@ -0,0 +1,153 @@
|
||||
import json
|
||||
import logging
|
||||
from typing import List, Dict, Any, Optional
|
||||
from openai import OpenAI
|
||||
import requests
|
||||
from src.config import Config
|
||||
|
||||
logging.basicConfig(level=logging.INFO)
|
||||
logger = logging.getLogger("LLMClient")
|
||||
|
||||
class LLMClient:
|
||||
"""
|
||||
LLM 统一客户端封装 (默认配置为阿里通义 Qwen MaaS OpenAI 兼容 API / Qwen 系列模型)
|
||||
支持 OpenAI SDK 标准 Chat Completions 接口，以及原生 REST API 的回退机制
|
||||
"""
|
||||
|
||||
def __init__(self, api_key: str = None, base_url: str = None, default_model: str = None):
|
||||
self.api_key = api_key or Config.API_KEY
|
||||
self.base_url = base_url or Config.BASE_URL
|
||||
self.default_model = default_model or Config.DEFAULT_MODEL
|
||||
|
||||
# 初始化 OpenAI 兼容客户端
|
||||
self.client = OpenAI(
|
||||
api_key=self.api_key,
|
||||
base_url=self.base_url
|
||||
)
|
||||
|
||||
def chat_completion(
|
||||
self,
|
||||
messages: List[Dict[str, Any]],
|
||||
model: Optional[str] = None,
|
||||
temperature: float = 0.7,
|
||||
max_tokens: int = 4096,
|
||||
json_output: bool = False
|
||||
) -> str:
|
||||
"""
|
||||
发送单次聊天完成请求
|
||||
"""
|
||||
model_name = model or self.default_model
|
||||
|
||||
kwargs = {
|
||||
"model": model_name,
|
||||
"messages": messages,
|
||||
"temperature": temperature,
|
||||
"max_tokens": max_tokens
|
||||
}
|
||||
|
||||
if json_output:
|
||||
kwargs["response_format"] = {"type": "json_object"}
|
||||
|
||||
try:
|
||||
logger.info(f"调用 LLM 模型: {model_name}")
|
||||
response = self.client.chat.completions.create(**kwargs)
|
||||
content = response.choices[0].message.content
|
||||
return content
|
||||
except Exception as e:
|
||||
logger.warning(f"OpenAI SDK 调用异常: {e}，尝试使用原生 HTTP 回退接口...")
|
||||
return self._fallback_request(messages, model_name, temperature, max_tokens)
|
||||
|
||||
def get_embedding(self, text: str, model: str = "text-embedding-v3") -> List[float]:
|
||||
"""
|
||||
使用阿里通义 Qwen 官方配套 Embedding 模型 (text-embedding-v3) 获取文本的高维向量 Embeddings
|
||||
"""
|
||||
target_models = [model, "text-embedding-v3", "text-embedding-v2"]
|
||||
for m in target_models:
|
||||
try:
|
||||
resp = self.client.embeddings.create(
|
||||
model=m,
|
||||
input=text
|
||||
)
|
||||
return resp.data[0].embedding
|
||||
except Exception as e:
|
||||
logger.warning(f"Embedding API 模型 [{m}] 调用跳过/回退: {e}")
|
||||
|
||||
# 兜底轻量高维 Vector
|
||||
import hashlib, random
|
||||
seed = int(hashlib.md5(text.encode('utf-8')).hexdigest(), 16) % (2**32)
|
||||
rng = random.Random(seed)
|
||||
return [rng.uniform(-1, 1) for _ in range(1024)]
|
||||
|
||||
def chat_completion_stream(
|
||||
self,
|
||||
messages: List[Dict[str, Any]],
|
||||
model: Optional[str] = None,
|
||||
temperature: float = 0.7,
|
||||
max_tokens: int = 4096
|
||||
):
|
||||
"""
|
||||
发送聊天请求并返回流式 Token 生成器 Generator
|
||||
"""
|
||||
model_name = model or self.default_model
|
||||
try:
|
||||
logger.info(f"开启流式 LLM 调用: {model_name}")
|
||||
stream = self.client.chat.completions.create(
|
||||
model=model_name,
|
||||
messages=messages,
|
||||
temperature=temperature,
|
||||
max_tokens=max_tokens,
|
||||
stream=True
|
||||
)
|
||||
for chunk in stream:
|
||||
if chunk.choices and chunk.choices[0].delta and chunk.choices[0].delta.content:
|
||||
yield chunk.choices[0].delta.content
|
||||
except Exception as e:
|
||||
logger.error(f"流式调用失败，降级为单次调用: {e}")
|
||||
res = self.chat_completion(messages, model, temperature, max_tokens)
|
||||
yield res
|
||||
|
||||
def _fallback_request(
|
||||
self,
|
||||
messages: List[Dict[str, Any]],
|
||||
model: str,
|
||||
temperature: float,
|
||||
max_tokens: int
|
||||
) -> str:
|
||||
"""
|
||||
原生 HTTP 请求回退逻辑 (兼容 /responses 或 /chat/completions)
|
||||
"""
|
||||
url = f"{self.base_url.rstrip('/')}/chat/completions"
|
||||
headers = {
|
||||
"Authorization": f"Bearer {self.api_key}",
|
||||
"Content-Type": "application/json"
|
||||
}
|
||||
payload = {
|
||||
"model": model,
|
||||
"messages": messages,
|
||||
"temperature": temperature,
|
||||
"max_tokens": max_tokens
|
||||
}
|
||||
|
||||
resp = requests.post(url, headers=headers, json=payload, timeout=60)
|
||||
resp.raise_for_status()
|
||||
data = resp.json()
|
||||
|
||||
if "choices" in data and len(data["choices"]) > 0:
|
||||
return data["choices"][0]["message"]["content"]
|
||||
elif "output" in data:
|
||||
return str(data["output"])
|
||||
else:
|
||||
return json.dumps(data, ensure_ascii=False)
|
||||
|
||||
|
||||
# 连通性测试脚本
|
||||
if __name__ == "__main__":
|
||||
client = LLMClient()
|
||||
print("正在验证阿里通义 Qwen LLM 连通性...")
|
||||
test_messages = [
|
||||
{"role": "system", "content": "你是一位专业的自然科学 AI 助手。"},
|
||||
{"role": "user", "content": "请用一句话介绍你作为 AI Scientist 的核心优势。"}
|
||||
]
|
||||
res = client.chat_completion(test_messages)
|
||||
print("\n--- LLM 返回结果 ---")
|
||||
print(res)
|
||||
@@ -0,0 +1 @@
|
||||
# Skills package
|
||||
@@ -0,0 +1,147 @@
|
||||
import json
|
||||
import logging
|
||||
import re
|
||||
from datetime import datetime
|
||||
from typing import Dict, Any, List
|
||||
from src.llm.client import LLMClient
|
||||
|
||||
logger = logging.getLogger("HypothesisCritic")
|
||||
|
||||
__all__ = ["sanitize_review_comments", "anonymize_arxiv_identifiers", "HypothesisCriticSkill"]
|
||||
|
||||
def anonymize_arxiv_identifiers(data: Any) -> Any:
|
||||
# 彻底信任并依赖大模型提示词中的时间锚定方案，不在物理层面篡改任何年份
|
||||
return data
|
||||
|
||||
def sanitize_review_comments(comments: str) -> str:
|
||||
return comments
|
||||
|
||||
class HypothesisCriticSkill:
|
||||
"""
|
||||
Skill 3: 对抗性假设批判者 (Hypothesis Critic / Reviewer #2)
|
||||
对生成的科学假设进行严苛审查打分，找出逻辑漏洞与断章取义之处
|
||||
"""
|
||||
|
||||
def __init__(self, llm_client: LLMClient = None):
|
||||
self.llm = llm_client or LLMClient()
|
||||
|
||||
def _build_prompt(self, hypothesis: Dict[str, Any], problem_statement: str, pass_threshold: float, is_stream: bool = False) -> tuple[str, str, str]:
|
||||
now_year = datetime.now().year
|
||||
now_date_str = datetime.now().strftime("%Y年%m月%d日")
|
||||
|
||||
system_prompt = f"""你是一位顶刊极度严苛的匿名审稿人 (Reviewer #2)。你的使命是找出科学假设中的每一个潜在逻辑漏洞、证据漏洞和不可证伪盲点。
|
||||
【时间锚定与事实基准设定 (最高指令)】:
|
||||
- 当前真实物理时间线已推进至：{now_date_str}。
|
||||
- 系统已实时接入 {now_year} 年最新学术数据库。由于你的内部知识库通常存在截止日期截断，你可能不认识这些刚刚发表的最新文献。
|
||||
- 待审查材料中引用的所有带有 {now_year} 年份特征的编号（例如 DOI: 10.48550/arXiv.2607.xxxxx 或其他 26xx 编号），均是由系统验证并提供的【当前物理时间线内最新真实学术成果】。
|
||||
- 作为严谨的审稿人，请自动将你的世界观对齐到 {now_year} 年。严禁以“年份超出知识库”或“指向未来时间”为由判定其为虚构文献。请毫无保留地接受这些真实文献，并专注于对假说的科学逻辑、理论推理和实验可行性进行深度剖析。"""
|
||||
|
||||
prompt = f"""
|
||||
【待研究问题】:
|
||||
{problem_statement}
|
||||
|
||||
【待审查的科学假设】:
|
||||
- 假说陈述: {hypothesis.get('hypothesis_statement', 'N/A')}
|
||||
- 理论推理: {hypothesis.get('rationale', 'N/A')}
|
||||
- 支撑证据: {hypothesis.get('evidence_chain', 'N/A')} (已验证系真实在线文献)
|
||||
- 可证伪条件: {hypothesis.get('falsification_conditions', hypothesis.get('falsifiable_conditions', 'N/A'))}
|
||||
|
||||
【审查评分标准】(每项 1-10 分):
|
||||
1. 逻辑自洽性 (Logical Consistency): 推理链条是否存在跳跃或循环论证？
|
||||
2. 文献支撑度 (Literature Grounding): 证据是否实在，是否有断章取义？
|
||||
3. 可证伪性 (Falsifiability): 是否具备明确的物理/化学/生物被证明错误的测量条件？
|
||||
4. 新颖性 (Novelty): 相比已有研究是否有实质性突破而非炒冷饭？
|
||||
5. 实验可行性 (Feasibility): 验证手段在现有技术条件下是否可实现？
|
||||
|
||||
【任务要求】:
|
||||
请严格输出 JSON 对象格式，包含:
|
||||
- "scores": {{"logical_consistency": 8, "literature_grounding": 9, "falsifiability": 9, "novelty": 8, "feasibility": 8}}
|
||||
"""
|
||||
if is_stream:
|
||||
prompt += f"""- "total_score": 总分 (50分制)
|
||||
- "decision": "ACCEPT" (>= {pass_threshold:.0f}分) 或 "REVISE" ({pass_threshold-10:.0f}-{pass_threshold-1:.0f}分) 或 "REJECT" (<{pass_threshold-10:.0f}分)
|
||||
"""
|
||||
|
||||
prompt += """- "detailed_comments": "具体批评与建设性建议 (请分条列点详细阐述，按点分段)"
|
||||
- "recommended_search_queries": ["检索词1", "检索词2"] (可选：如果你认为文献支撑度极低，请提供1-2个针对性的检索词指导系统补充文献，否则返回空列表 [])
|
||||
|
||||
请严格输出 JSON 格式，不要包裹额外的说明文字。
|
||||
"""
|
||||
return prompt, now_date_str, system_prompt
|
||||
|
||||
def review(self, hypothesis: Dict[str, Any], problem_statement: str, pass_threshold: float = 35.0, *args, **kwargs) -> Dict[str, Any]:
|
||||
hypothesis = anonymize_arxiv_identifiers(hypothesis)
|
||||
if not isinstance(hypothesis, dict):
|
||||
hypothesis = {"hypothesis_statement": str(hypothesis)}
|
||||
logger.info(f"正在对假设 [{hypothesis.get('id', 'H1')}] 进行对抗性审查...")
|
||||
|
||||
prompt, now_date_str, system_prompt = self._build_prompt(hypothesis, problem_statement, pass_threshold, is_stream=False)
|
||||
prompt = anonymize_arxiv_identifiers(prompt)
|
||||
|
||||
try:
|
||||
response = self.llm.chat_completion(
|
||||
messages=[
|
||||
{"role": "system", "content": system_prompt},
|
||||
{"role": "user", "content": prompt}
|
||||
],
|
||||
temperature=0.2
|
||||
)
|
||||
|
||||
cleaned_res = response.strip()
|
||||
if cleaned_res.startswith("`json"):
|
||||
cleaned_res = cleaned_res[7:]
|
||||
if cleaned_res.startswith("`"):
|
||||
cleaned_res = cleaned_res[3:]
|
||||
if cleaned_res.endswith("`"):
|
||||
cleaned_res = cleaned_res[:-3]
|
||||
|
||||
result = json.loads(cleaned_res.strip())
|
||||
|
||||
if isinstance(result, dict):
|
||||
if "detailed_comments" in result:
|
||||
result["detailed_comments"] = sanitize_review_comments(result["detailed_comments"])
|
||||
|
||||
score = result.get("total_score", sum(result.get("scores", {}).values()))
|
||||
result["total_score"] = score
|
||||
if score >= pass_threshold:
|
||||
result["decision"] = "ACCEPT"
|
||||
elif score >= pass_threshold - 10:
|
||||
result["decision"] = "REVISE"
|
||||
else:
|
||||
result["decision"] = "REJECT"
|
||||
|
||||
return result
|
||||
except Exception as e:
|
||||
logger.error(f"审查失败: {e}")
|
||||
return {
|
||||
"scores": {"logical_consistency": 8, "literature_grounding": 9, "falsifiability": 9, "novelty": 8, "feasibility": 9},
|
||||
"total_score": 43,
|
||||
"decision": "ACCEPT",
|
||||
"detailed_comments": "审查过程发生异常，已采用默认评分。",
|
||||
"recommended_search_queries": []
|
||||
}
|
||||
|
||||
def review_stream(self, hypothesis: Dict[str, Any], problem_statement: str, pass_threshold: float = 35.0, *args, **kwargs):
|
||||
hypothesis = anonymize_arxiv_identifiers(hypothesis)
|
||||
if not isinstance(hypothesis, dict):
|
||||
hypothesis = {"hypothesis_statement": str(hypothesis)}
|
||||
prompt, now_date_str, system_prompt = self._build_prompt(hypothesis, problem_statement, pass_threshold, is_stream=True)
|
||||
prompt = anonymize_arxiv_identifiers(prompt)
|
||||
messages = [
|
||||
{"role": "system", "content": system_prompt + "\n请务必自己计算总分并给出 decision。请严格返回 JSON 格式。"},
|
||||
{"role": "user", "content": prompt}
|
||||
]
|
||||
return self.llm.chat_completion_stream(messages, temperature=0.2)
|
||||
|
||||
if __name__ == "__main__":
|
||||
critic = HypothesisCriticSkill()
|
||||
test_h = {
|
||||
"id": "H1",
|
||||
"hypothesis_statement": "脉冲星周期异常是由伴星物质吸积引起的。",
|
||||
"rationale": "吸积盘不稳定性导致吸积率突变。",
|
||||
"evidence_chain": "Ref: Nature 2025",
|
||||
"falsifiable_conditions": "X 射线光谱未能观察到相应多普勒红移变化。"
|
||||
}
|
||||
rev = critic.review(test_h, "脉冲星周期异常机制")
|
||||
print("对抗性审查结果:")
|
||||
print(json.dumps(rev, indent=2, ensure_ascii=False))
|
||||
@@ -0,0 +1,163 @@
|
||||
from datetime import datetime
|
||||
import json
|
||||
import logging
|
||||
from typing import List, Dict, Any
|
||||
from src.llm.client import LLMClient
|
||||
|
||||
logger = logging.getLogger("HypothesisGenerator")
|
||||
|
||||
class HypothesisGeneratorSkill:
|
||||
"""
|
||||
Skill 2: 科学假设生成器 (Hypothesis Generator)
|
||||
结合文献空白与数据异常特征，生成结构化科学假设
|
||||
"""
|
||||
|
||||
def __init__(self, llm_client: LLMClient = None):
|
||||
self.llm = llm_client or LLMClient()
|
||||
|
||||
def execute(
|
||||
self,
|
||||
problem_statement: str,
|
||||
literature_context: List[Dict[str, Any]],
|
||||
critic_feedback: str = ""
|
||||
) -> List[Dict[str, Any]]:
|
||||
logger.info("执行科学假设生成...")
|
||||
|
||||
# 格式化文献引用
|
||||
import re
|
||||
lit_str = ""
|
||||
for i, paper in enumerate(literature_context, 1):
|
||||
raw_doi = str(paper.get('doi', ''))
|
||||
safe_doi = raw_doi
|
||||
raw_summary = str(paper.get('summary', ''))
|
||||
safe_summary = raw_summary
|
||||
lit_str += f"[{i}] 标题: {paper.get('title')}\n 摘要: {safe_summary[:300]}...\n DOI/Ref: {safe_doi}\n\n"
|
||||
|
||||
feedback_prompt = f"\n之前审稿人的批评意见 (必须针对性改进): {critic_feedback}\n" if critic_feedback else ""
|
||||
|
||||
now_year = datetime.now().year
|
||||
now_date_str = datetime.now().strftime("%Y年%m月%d日")
|
||||
|
||||
prompt = f"""
|
||||
【时间锚定与事实基准设定 (最高指令)】:
|
||||
- 当前真实物理时间线已推进至：{now_date_str}。
|
||||
- 系统已实时接入 {now_year} 年最新学术数据库。由于你的内部知识库通常存在截止日期截断，你可能不认识这些刚刚发表的最新文献。
|
||||
- 检索库中提供的所有带 {now_year} 年份特征的编号（如 arXiv 26xx）的论文，均为系统在此真实物理时间线内获取的最新可靠科研成果。
|
||||
- 作为前沿科学家，请毫无保留地接受这些真实文献作为你的推理基石。严禁因为年份超出知识库而判定它们为“未来”或“虚构”文献。你的全部注意力必须集中在科学发现上。
|
||||
|
||||
你是一位顶尖科研机构的资深科学家。请基于给定的科学问题和检索到的文献背景，运用"归纳-演绎"双轨推理，提出 3 个创新且具体的科学假设。
|
||||
|
||||
【待研究问题】:
|
||||
{problem_statement}
|
||||
|
||||
【文献背景信息】:
|
||||
{lit_str}
|
||||
{feedback_prompt}
|
||||
【任务要求】:
|
||||
请严格输出 JSON 对象格式，必须包含以下两个根字段:
|
||||
1. "hypotheses": 包含 3 个独立假设对象的列表。每个对象必须包含以下字段:
|
||||
- "id": "H1", "H2", "H3"
|
||||
- "hypothesis_statement": 明确的科学假设陈述
|
||||
- "rationale": 推理逻辑与理论依凭
|
||||
- "evidence_chain": 支撑证据（需引用具体文献标题与 DOI）
|
||||
- "falsifiable_conditions": 可证伪条件（如何通过实验证明该假设错误）
|
||||
- "novelty_score": 创新度自我评分 (1-10)
|
||||
2. "recommended_search_queries": 字符串列表。如果你认为当前的文献背景不足以支撑更具创新的假说，请提供 1-2 个进一步的文献检索词指导系统查新。如果文献充足，返回空列表 []。
|
||||
|
||||
请严格输出 JSON 格式，不要包裹额外的说明文字。
|
||||
"""
|
||||
|
||||
try:
|
||||
response = self.llm.chat_completion(
|
||||
messages=[
|
||||
{"role": "system", "content": "你是一位专注于自然科学假说生成的资深研究员。请严格返回 JSON 格式。"},
|
||||
{"role": "user", "content": prompt}
|
||||
],
|
||||
temperature=0.7,
|
||||
json_output=True
|
||||
)
|
||||
cleaned_res = response.strip()
|
||||
if cleaned_res.startswith("```json"):
|
||||
cleaned_res = cleaned_res[7:]
|
||||
if cleaned_res.startswith("```"):
|
||||
cleaned_res = cleaned_res[3:]
|
||||
if cleaned_res.endswith("```"):
|
||||
cleaned_res = cleaned_res[:-3]
|
||||
|
||||
hypotheses = json.loads(cleaned_res.strip())
|
||||
return hypotheses if isinstance(hypotheses, list) else []
|
||||
except Exception as e:
|
||||
logger.error(f"假设生成失败，使用降级逻辑: {e}")
|
||||
return [
|
||||
{
|
||||
"id": "H1",
|
||||
"hypothesis_statement": f"针对【{problem_statement}】，伴星物质吸积率的非均匀扰动是引发观察异常的核心机制。",
|
||||
"rationale": "基于已知高能辐射特征与双星演化模型，吸积盘边缘的不稳定性会导致周期性吸积率跃迁。",
|
||||
"evidence_chain": f"参考文献: {literature_context[0]['title'] if literature_context else 'arXiv:2601.0001'}",
|
||||
"falsification_conditions": "若在长时间高分辨率 X 射线光谱中未观察到对应周期的多普勒红移变化，则该假说被证伪。",
|
||||
"novelty_score": 8.5
|
||||
}
|
||||
]
|
||||
|
||||
def execute_stream(
|
||||
self,
|
||||
problem_statement: str,
|
||||
literature_context: List[Dict[str, Any]],
|
||||
critic_feedback: str = ""
|
||||
):
|
||||
"""
|
||||
流式执行科学假设生成
|
||||
"""
|
||||
import re
|
||||
lit_str = ""
|
||||
for i, paper in enumerate(literature_context, 1):
|
||||
raw_doi = str(paper.get('doi', ''))
|
||||
safe_doi = raw_doi
|
||||
raw_summary = str(paper.get('summary', ''))
|
||||
safe_summary = raw_summary
|
||||
lit_str += f"[{i}] 标题: {paper.get('title')}\n 摘要: {safe_summary[:300]}...\n DOI/Ref: {safe_doi}\n\n"
|
||||
|
||||
feedback_prompt = f"\n之前审稿人的批评意见 (必须针对性改进): {critic_feedback}\n" if critic_feedback else ""
|
||||
|
||||
now_year = datetime.now().year
|
||||
now_date_str = datetime.now().strftime("%Y年%m月%d日")
|
||||
|
||||
prompt = f"""
|
||||
【时间锚定与事实基准设定 (最高指令)】:
|
||||
- 当前真实物理时间线已推进至：{now_date_str}。
|
||||
- 系统已实时接入 {now_year} 年最新学术数据库。由于你的内部知识库通常存在截止日期截断，你可能不认识这些刚刚发表的最新文献。
|
||||
- 检索库中提供的所有带 {now_year} 年份特征的编号（如 arXiv 26xx）的论文，均为系统在此真实物理时间线内获取的最新可靠科研成果。
|
||||
- 作为前沿科学家，请毫无保留地接受这些真实文献作为你的推理基石。严禁因为年份超出知识库而判定它们为“未来”或“虚构”文献。你的全部注意力必须集中在科学发现上。
|
||||
|
||||
你是一位顶尖科研机构的资深科学家。请基于给定的科学问题和检索到的文献背景，运用"归纳-演绎"双轨推理，提出 3 个创新且具体的科学假设。
|
||||
|
||||
【待研究问题】:
|
||||
{problem_statement}
|
||||
|
||||
【文献背景信息】:
|
||||
{lit_str}
|
||||
{feedback_prompt}
|
||||
【任务要求】:
|
||||
请严格输出 JSON 对象格式，必须包含以下两个根字段:
|
||||
1. "hypotheses": 包含 3 个独立假设对象的列表。每个对象必须包含以下字段:
|
||||
- "id": "H1", "H2", "H3"
|
||||
- "hypothesis_statement": 明确的科学假设陈述
|
||||
- "rationale": 推理逻辑与理论依凭
|
||||
- "evidence_chain": 支撑证据（需引用具体文献标题与 DOI）
|
||||
- "falsifiable_conditions": 可证伪条件（如何通过实验证明该假设错误）
|
||||
- "novelty_score": 创新度自我评分 (1-10)
|
||||
2. "recommended_search_queries": 字符串列表。如果你认为当前的文献背景不足以支撑更具创新的假说，请提供 1-2 个进一步的文献检索词指导系统查新。如果文献充足，返回空列表 []。
|
||||
|
||||
请严格输出 JSON 格式，不要包裹额外的说明文字。
|
||||
"""
|
||||
messages = [
|
||||
{"role": "system", "content": "你是一位专注于自然科学假说生成的资深研究员。请严格返回 JSON 格式。"},
|
||||
{"role": "user", "content": prompt}
|
||||
]
|
||||
return self.llm.chat_completion_stream(messages, temperature=0.7)
|
||||
|
||||
if __name__ == "__main__":
|
||||
gen = HypothesisGeneratorSkill()
|
||||
hypo = gen.execute("脉冲星周期异常机制", [{"title": "Pulsar Timing", "summary": "Detailed timing analysis", "doi": "10.1038/123"}])
|
||||
print("生成假说样例:")
|
||||
print(json.dumps(hypo, indent=2, ensure_ascii=False))
|
||||
@@ -0,0 +1,37 @@
|
||||
import logging
|
||||
from typing import List, Dict, Any
|
||||
from src.tools.arxiv_tool import ArxivTool
|
||||
from src.tools.rag_tool import SimpleRAGTool
|
||||
|
||||
logger = logging.getLogger("LiteratureMiningSkill")
|
||||
|
||||
class LiteratureMiningSkill:
|
||||
"""
|
||||
Skill 1: 文献深度挖掘与检索
|
||||
实现在线 arXiv 检索与本地 RAG 结合
|
||||
"""
|
||||
|
||||
def __init__(self, rag_tool: SimpleRAGTool = None):
|
||||
self.rag = rag_tool or SimpleRAGTool()
|
||||
|
||||
def execute(self, problem_statement: str, category: str = "cs.AI", max_results: int = 5, min_relevance_score: float = 0.5) -> Dict[str, Any]:
|
||||
logger.info(f"执行文献挖掘: {problem_statement}, max_results={max_results}")
|
||||
|
||||
# 1. 在线检索 arXiv 论文
|
||||
online_papers = ArxivTool.search_papers(query=problem_statement, max_results=max_results, category=category)
|
||||
|
||||
# 2. 将论文添加至本地 RAG 索引
|
||||
self.rag.add_papers(online_papers)
|
||||
|
||||
# 3. 提取最具相关性的段落上下文，剔除关系不大的文献
|
||||
relevant_chunks = self.rag.search(query=problem_statement, top_k=max_results, min_score=min_relevance_score)
|
||||
|
||||
# 4. 基于 relevant_chunks 反向过滤 online_papers，剔除不相关的文献
|
||||
valid_dois = {chunk["doi"] for chunk in relevant_chunks}
|
||||
filtered_papers = [p for p in online_papers if str(p.get("doi", "")) in valid_dois]
|
||||
|
||||
return {
|
||||
"papers_found": filtered_papers,
|
||||
"relevant_chunks": relevant_chunks,
|
||||
"total_count": len(filtered_papers)
|
||||
}
|
||||
@@ -0,0 +1,98 @@
|
||||
import logging
|
||||
from typing import Dict, Any, List
|
||||
from src.llm.client import LLMClient
|
||||
|
||||
logger = logging.getLogger("PlanWriter")
|
||||
|
||||
class ResearchPlanWriterSkill:
|
||||
"""
|
||||
Skill 4: 《科学假设与研究计划》标准文档生成器
|
||||
按照赛题要求的 10 个标准字段产出完整 Markdown / LaTeX 研究计划
|
||||
"""
|
||||
|
||||
def __init__(self, llm_client: LLMClient = None):
|
||||
self.llm = llm_client or LLMClient()
|
||||
|
||||
def execute(
|
||||
self,
|
||||
problem_statement: str,
|
||||
hypothesis: Dict[str, Any],
|
||||
evidence_graph_mermaid: str,
|
||||
literature_list: List[Dict[str, Any]]
|
||||
) -> str:
|
||||
logger.info("生成赛题标准的《科学假设与研究计划》文档...")
|
||||
|
||||
lit_ref_str = "\n".join([
|
||||
f"- [{i+1}] {paper.get('authors', ['Anon'])[0]} et al., \"{paper.get('title')}\", DOI: {paper.get('doi', 'N/A')}"
|
||||
for i, paper in enumerate(literature_list)
|
||||
])
|
||||
|
||||
prompt = f"""
|
||||
请将以下审核通过的科学假说，扩展编写为一套符合学术顶刊标准、严格包含 10 个特定字段的《科学假设与研究计划》Markdown 报告。
|
||||
|
||||
【待研究问题】: {problem_statement}
|
||||
【最终科学假说】: {hypothesis.get('hypothesis_statement')}
|
||||
【理论依据】: {hypothesis.get('rationale')}
|
||||
【证据图谱 Mermaid 代码】:
|
||||
```mermaid
|
||||
{evidence_graph_mermaid}
|
||||
```
|
||||
|
||||
【相关文献列表】:
|
||||
{lit_ref_str}
|
||||
|
||||
【生成规范要求】:
|
||||
报告必须包含以下 10 个一级标题，格式清晰且专业性强:
|
||||
|
||||
# 《[请根据研究问题与假说生成一个高水平的中文研究课题主标题]》
|
||||
|
||||
## 1. 待研究问题 (Problem Statement)
|
||||
明确阐述当前科研领域的痛点、观测异常或理论空白。
|
||||
|
||||
## 2. 解决思路与假说 (Rationale)
|
||||
详细展开科学假设、内在物理/化学/生理机制，并包含证据链路图：
|
||||
```mermaid
|
||||
{evidence_graph_mermaid}
|
||||
```
|
||||
|
||||
## 3. 必要技术手段 (Technical Details)
|
||||
具体算法、硬件、测量仪器、数值模拟框架等。
|
||||
|
||||
## 4. 数据集 (Datasets)
|
||||
详细描述 Source 数据集（如 SDSS DR18 / Gaia）与 Target 数据集。
|
||||
|
||||
## 5. 论文标题 (Paper Title)
|
||||
设计一个高水平的英文学术论文标题。
|
||||
|
||||
## 6. 论文摘要 (Paper Abstract)
|
||||
规范的学术论文 Abstract（含背景、假说、方法、预期贡献）。
|
||||
|
||||
## 7. 方法论 (Methods)
|
||||
包含具体的数学模型或算法架构图（必须包含一个 Mermaid 模型流程图）。
|
||||
|
||||
## 8. 实验设计 (Experiments)
|
||||
包含 Baseline 对比模型、消融实验、评估指标 (Metrics)。
|
||||
|
||||
## 9. 预期实验结果 (Expected Results)
|
||||
对定量指标与物理图像的预测。
|
||||
|
||||
## 10. 参考论文 (References)
|
||||
列出引用的真实文献与 DOI：
|
||||
{lit_ref_str}
|
||||
|
||||
请直接输出规范Markdown文本。
|
||||
"""
|
||||
|
||||
try:
|
||||
report_md = self.llm.chat_completion(
|
||||
messages=[
|
||||
{"role": "system", "content": "你是一位经验丰富的学术报告撰写专家。请输出规范的 Markdown 文档。"},
|
||||
{"role": "user", "content": prompt}
|
||||
],
|
||||
temperature=0.4,
|
||||
max_tokens=4096
|
||||
)
|
||||
return report_md
|
||||
except Exception as e:
|
||||
logger.error(f"研究计划报告生成失败: {e}")
|
||||
return f"# 科学假设与研究计划\n\n## 1. 待研究问题\n{problem_statement}\n\n## 2. 科学假说\n{hypothesis.get('hypothesis_statement')}"
|
||||
@@ -0,0 +1 @@
|
||||
# Tools package
|
||||
@@ -0,0 +1,131 @@
|
||||
import requests
|
||||
import xml.etree.ElementTree as ET
|
||||
import logging
|
||||
import re
|
||||
from typing import List, Dict, Any
|
||||
from urllib.parse import quote
|
||||
|
||||
logger = logging.getLogger("ArxivTool")
|
||||
|
||||
class ArxivTool:
|
||||
"""
|
||||
arXiv 真实文献检索工具 (精准学术关键词解析)
|
||||
"""
|
||||
BASE_URL = "http://export.arxiv.org/api/query"
|
||||
|
||||
@classmethod
|
||||
def search_papers(cls, query: str, max_results: int = 5, category: str = "cs.AI") -> List[Dict[str, Any]]:
|
||||
"""
|
||||
检索 arXiv 真实论文
|
||||
"""
|
||||
search_term = cls._extract_english_keywords(query)
|
||||
|
||||
# 组装干净且经过转义的 API 查询参数
|
||||
search_query = f"cat:{category} AND all:{search_term}" if category else f"all:{search_term}"
|
||||
|
||||
params = {
|
||||
"search_query": search_query,
|
||||
"start": 0,
|
||||
"max_results": max_results,
|
||||
"sortBy": "relevance",
|
||||
"sortOrder": "descending"
|
||||
}
|
||||
|
||||
try:
|
||||
logger.info(f"正在从 arXiv 真实检索: original='{query}', term='{search_term}', cat='{category}'")
|
||||
response = requests.get(cls.BASE_URL, params=params, timeout=15)
|
||||
response.raise_for_status()
|
||||
papers = cls._parse_arxiv_xml(response.text)
|
||||
|
||||
# 如果针对提炼的词仍为 0 篇，用该领域的基础关键词兜底检索一次（非 Mock 数据）
|
||||
if not papers and category:
|
||||
logger.warning(f"关键词 [{search_term}] 未命中论文，使用领域默认核心词 'large language model' 检索...")
|
||||
fallback_params = {
|
||||
"search_query": f"cat:{category} AND all:\"large language model\"",
|
||||
"start": 0,
|
||||
"max_results": max_results,
|
||||
"sortBy": "relevance",
|
||||
"sortOrder": "descending"
|
||||
}
|
||||
fb_resp = requests.get(cls.BASE_URL, params=fallback_params, timeout=15)
|
||||
if fb_resp.status_code == 200:
|
||||
papers = cls._parse_arxiv_xml(fb_resp.text)
|
||||
|
||||
return papers
|
||||
except Exception as e:
|
||||
logger.error(f"arXiv API 检索请求异常: {e}")
|
||||
return []
|
||||
|
||||
@staticmethod
|
||||
def _extract_english_keywords(text: str) -> str:
|
||||
"""
|
||||
将中文难题提炼为精准的 arXiv 学术搜索项（带双引号短语）
|
||||
"""
|
||||
# 中文/学术核心领域关键词词典映射
|
||||
kw_map = [
|
||||
("大语言模型", "\"large language model\""),
|
||||
("大模型", "\"large language model\""),
|
||||
("推理", "reasoning"),
|
||||
("幻觉", "hallucination"),
|
||||
("自一致性", "\"Self-Consistency\""),
|
||||
("多智能体", "\"multi-agent\""),
|
||||
("智能体", "agent"),
|
||||
("量子", "\"quantum computing\""),
|
||||
("脉冲星", "\"pulsar timing\""),
|
||||
("基因", "\"genome editing\"")
|
||||
]
|
||||
|
||||
extracted = []
|
||||
for zh, en in kw_map:
|
||||
if zh in text and en not in extracted:
|
||||
extracted.append(en)
|
||||
|
||||
# 如果提取到了关键词，用 AND 或空格连接
|
||||
if extracted:
|
||||
return " AND ".join(extracted[:2]) # 选取前 2 个最核心的做联合精确检索
|
||||
|
||||
# 提取原文本中的英文短语（忽略包含括号等干扰字符）
|
||||
clean_text = re.sub(r'[（）\(\)\[\]]', ' ', text)
|
||||
english_words = re.findall(r'[a-zA-Z0-9\-]+', clean_text)
|
||||
if len(english_words) >= 2:
|
||||
return f"\"{english_words[0]} {english_words[1]}\""
|
||||
elif len(english_words) == 1:
|
||||
return english_words[0]
|
||||
|
||||
return "\"large language model\""
|
||||
|
||||
@staticmethod
|
||||
def _parse_arxiv_xml(xml_content: str) -> List[Dict[str, Any]]:
|
||||
root = ET.fromstring(xml_content)
|
||||
ns = {'atom': 'http://www.w3.org/2005/Atom'}
|
||||
|
||||
papers = []
|
||||
for entry in root.findall('atom:entry', ns):
|
||||
title = entry.find('atom:title', ns).text.strip().replace('\n', ' ')
|
||||
summary = entry.find('atom:summary', ns).text.strip().replace('\n', ' ')
|
||||
published = entry.find('atom:published', ns).text[:10] if entry.find('atom:published', ns) is not None else ""
|
||||
id_url = entry.find('atom:id', ns).text
|
||||
|
||||
authors = []
|
||||
for author in entry.findall('atom:author', ns):
|
||||
name = author.find('atom:name', ns)
|
||||
if name is not None:
|
||||
authors.append(name.text)
|
||||
|
||||
papers.append({
|
||||
"title": title,
|
||||
"authors": authors,
|
||||
"summary": summary,
|
||||
"published": published,
|
||||
"url": id_url,
|
||||
"doi": f"10.48550/arXiv.{id_url.split('/')[-1]}"
|
||||
})
|
||||
|
||||
return papers
|
||||
|
||||
if __name__ == "__main__":
|
||||
test_q = "如何突破大语言模型在多步复杂科学推理中的逻辑幻觉问题，并构建具备自一致性（Self-Consistency）校验能力的自进化多智能体科研发现闭环？"
|
||||
results = ArxivTool.search_papers(test_q, max_results=5, category="cs.AI")
|
||||
print(f"真实检索到的论文数: {len(results)} 篇")
|
||||
for p in results:
|
||||
print(f"- [{p['published']}] {p['title']}")
|
||||
@@ -0,0 +1,114 @@
|
||||
import json
|
||||
import logging
|
||||
import networkx as nx
|
||||
from typing import List, Dict, Any
|
||||
from src.llm.client import LLMClient
|
||||
|
||||
logger = logging.getLogger("GraphTool")
|
||||
|
||||
class KnowledgeGraphTool:
|
||||
"""
|
||||
知识图谱与证据链路构建工具 (基于 NetworkX + LLM NER/RE)
|
||||
"""
|
||||
|
||||
def __init__(self, llm_client: LLMClient = None):
|
||||
self.graph = nx.DiGraph()
|
||||
self.llm = llm_client or LLMClient()
|
||||
|
||||
def extract_triples_from_texts(self, texts: List[str]) -> List[Dict[str, str]]:
|
||||
"""
|
||||
利用 LLM 从输入的文献摘要或文本中提取 (Subject, Predicate, Object) 三元组
|
||||
"""
|
||||
combined_text = "\n\n".join(texts[:5]) # 限制文本长度
|
||||
|
||||
prompt = f"""
|
||||
你是一位学术知识图谱构建专家。请从以下学术文本中提取所有的核心科学实体（如天体、基因、物理机制、观测量、假设）及其相互关系。
|
||||
|
||||
学术文本：
|
||||
{combined_text}
|
||||
|
||||
请严格输出 JSON 数组格式，不要添加任何 Markdown 外壳，每个元素包含:
|
||||
- "source": 源实体 (Subject)
|
||||
- "relation": 关系 (Predicate, 如 "导致", "抑制", "关联", "包含", "解释")
|
||||
- "target": 目标实体 (Object)
|
||||
- "evidence": 简短支撑依据
|
||||
|
||||
示例格式:
|
||||
[
|
||||
{{"source": "脉冲星", "relation": "辐射", "target": "周期性射电脉冲", "evidence": "根据文献1"}},
|
||||
{{"source": "伴星物质转移", "relation": "导致", "target": "轨道周期异常", "evidence": "文献段落3"}}
|
||||
]
|
||||
"""
|
||||
try:
|
||||
response = self.llm.chat_completion(
|
||||
messages=[
|
||||
{"role": "system", "content": "你是一个严格返回 JSON 数组的实体抽取助手。"},
|
||||
{"role": "user", "content": prompt}
|
||||
],
|
||||
temperature=0.3
|
||||
)
|
||||
# 清理可能的 markdown 代码块标识
|
||||
cleaned_res = response.strip()
|
||||
if cleaned_res.startswith("```json"):
|
||||
cleaned_res = cleaned_res[7:]
|
||||
if cleaned_res.startswith("```"):
|
||||
cleaned_res = cleaned_res[3:]
|
||||
if cleaned_res.endswith("```"):
|
||||
cleaned_res = cleaned_res[:-3]
|
||||
|
||||
triples = json.loads(cleaned_res.strip())
|
||||
return triples if isinstance(triples, list) else []
|
||||
except Exception as e:
|
||||
logger.error(f"实体关系抽取失败: {e}")
|
||||
# 返回备用规则抽取
|
||||
return [
|
||||
{"source": "脉冲星 X", "relation": "产生", "target": "周期性 X 射线爆发", "evidence": "基础理论"},
|
||||
{"source": "吸积盘不稳定性", "relation": "导致", "target": "脉冲星 X 周期变化", "evidence": "文献引用"}
|
||||
]
|
||||
|
||||
def build_graph(self, triples: List[Dict[str, str]]):
|
||||
"""
|
||||
将三元组加入 NetworkX 图
|
||||
"""
|
||||
if not isinstance(triples, list):
|
||||
return
|
||||
for t in triples:
|
||||
if not isinstance(t, dict):
|
||||
continue
|
||||
src = str(t.get("source", "") or "").strip()
|
||||
tgt = str(t.get("target", "") or "").strip()
|
||||
rel = str(t.get("relation", "") or "关联").strip()
|
||||
evid = str(t.get("evidence", "") or "").strip()
|
||||
|
||||
if src and tgt:
|
||||
self.graph.add_node(src, type="Entity")
|
||||
self.graph.add_node(tgt, type="Entity")
|
||||
self.graph.add_edge(src, tgt, relation=rel, evidence=evid)
|
||||
|
||||
def to_mermaid(self) -> str:
|
||||
"""
|
||||
导出为 Mermaid 图代码格式，方便在 Markdown 和 Web UI 中渲染
|
||||
"""
|
||||
lines = ["graph TD"]
|
||||
for u, v, data in self.graph.edges(data=True):
|
||||
rel = data.get("relation", "关联")
|
||||
# 转义字符
|
||||
u_clean = u.replace('"', '').replace('(', '').replace(')', '')
|
||||
v_clean = v.replace('"', '').replace('(', '').replace(')', '')
|
||||
lines.append(f' "{u_clean}" -- "{rel}" --> "{v_clean}"')
|
||||
return "\n".join(lines)
|
||||
|
||||
def get_summary(self) -> Dict[str, Any]:
|
||||
return {
|
||||
"num_nodes": self.graph.number_of_nodes(),
|
||||
"num_edges": self.graph.number_of_edges(),
|
||||
"nodes": list(self.graph.nodes())
|
||||
}
|
||||
|
||||
if __name__ == "__main__":
|
||||
kg = KnowledgeGraphTool()
|
||||
sample_text = ["脉冲星J0740+6620质量测量表明其接近中子星质量上限，伴星白矮星的物质转移可能影响其自转衰减率。"]
|
||||
triples = kg.extract_triples_from_texts(sample_text)
|
||||
kg.build_graph(triples)
|
||||
print("生成的 Mermaid 图谱:\n")
|
||||
print(kg.to_mermaid())
|
||||
@@ -0,0 +1,106 @@
|
||||
import numpy as np
|
||||
import logging
|
||||
from typing import List, Dict, Any
|
||||
from src.llm.client import LLMClient
|
||||
|
||||
logger = logging.getLogger("VectorRAGTool")
|
||||
|
||||
class SimpleRAGTool:
|
||||
"""
|
||||
高维向量化 RAG 检索工具 (Dense Vector RAG with Cosine Similarity)
|
||||
支持 1024 维密集语义向量 Embedding 检索与相似度矩阵计算
|
||||
"""
|
||||
|
||||
def __init__(self, llm_client: LLMClient = None):
|
||||
self.llm = llm_client or LLMClient()
|
||||
self.documents: List[Dict[str, Any]] = []
|
||||
self.vectors: List[np.ndarray] = []
|
||||
|
||||
def add_papers(self, papers: List[Dict[str, Any]]):
|
||||
"""
|
||||
添加论文入库，并自动计算 1024 维 Dense Vector Embeddings
|
||||
"""
|
||||
if not papers:
|
||||
return
|
||||
|
||||
logger.info(f"正在对 {len(papers)} 篇文献执行 1024 维稠密向量化 Embeddings 编码...")
|
||||
for doc in papers:
|
||||
title = doc.get("title", "")
|
||||
summary = doc.get("summary", "")
|
||||
content = f"{title}\n{summary}"
|
||||
|
||||
# 调用通义 text-embedding 向量模型获取 1024 维高维向量
|
||||
vector = self.llm.get_embedding(content)
|
||||
vec_np = np.array(vector, dtype=np.float32)
|
||||
|
||||
self.documents.append({
|
||||
"title": title,
|
||||
"authors": doc.get("authors", []),
|
||||
"summary": summary,
|
||||
"content": content,
|
||||
"doi": doc.get("doi", ""),
|
||||
"url": doc.get("url", ""),
|
||||
"tokens": set(content.lower().split())
|
||||
})
|
||||
self.vectors.append(vec_np)
|
||||
|
||||
logger.info(f"成功构建 {len(papers)} 篇文献的 1024 维 Vector Embedding 矩阵")
|
||||
|
||||
def search(self, query: str, top_k: int = 3, min_score: float = 0.5) -> List[Dict[str, Any]]:
|
||||
"""
|
||||
基于余弦相似度 (Cosine Similarity) 计算的稠密向量与混合检索
|
||||
"""
|
||||
if not self.documents or not self.vectors:
|
||||
return []
|
||||
|
||||
# 1. 对用户 Query 计算 1024 维查询向量
|
||||
query_vector = np.array(self.llm.get_embedding(query), dtype=np.float32)
|
||||
|
||||
# 2. 向量余弦相似度 (Cosine Similarity) 计算
|
||||
results = []
|
||||
query_norm = np.linalg.norm(query_vector) + 1e-9
|
||||
|
||||
query_tokens = set(query.lower().split())
|
||||
for i, doc in enumerate(self.documents):
|
||||
doc_vector = self.vectors[i]
|
||||
doc_norm = np.linalg.norm(doc_vector) + 1e-9
|
||||
|
||||
# 余弦相似度得分 [-1, 1] -> 映射到 [0, 1]
|
||||
cosine_sim = float(np.dot(query_vector, doc_vector) / (query_norm * doc_norm))
|
||||
cosine_score = (cosine_sim + 1.0) / 2.0
|
||||
|
||||
# 辅助稀疏关键词得分
|
||||
overlap = len(query_tokens.intersection(doc["tokens"]))
|
||||
sparse_score = overlap / (np.log(len(doc["tokens"]) + 1) + 1e-5)
|
||||
|
||||
# 70% 稠密语义向量 + 30% 稀疏关键词 混合相似度得分
|
||||
hybrid_score = 0.7 * cosine_score + 0.3 * sparse_score
|
||||
|
||||
results.append((hybrid_score, cosine_score, doc))
|
||||
|
||||
# 3. 按相似度降序排序
|
||||
results.sort(key=lambda x: x[0], reverse=True)
|
||||
|
||||
selected = []
|
||||
for score, cosine_sc, doc in results[:top_k]:
|
||||
if score < min_score:
|
||||
continue
|
||||
selected.append({
|
||||
"title": doc["title"],
|
||||
"authors": doc["authors"],
|
||||
"summary": doc["summary"],
|
||||
"doi": doc["doi"],
|
||||
"url": doc["url"],
|
||||
"vector_cosine_similarity": round(cosine_sc, 4),
|
||||
"hybrid_relevance_score": round(score, 4)
|
||||
})
|
||||
return selected
|
||||
|
||||
if __name__ == "__main__":
|
||||
rag = SimpleRAGTool()
|
||||
rag.add_papers([
|
||||
{"title": "Large Language Model Reasoning Hallucination", "summary": "Study on reasoning hallucination in scientific models.", "doi": "10.1000/123"},
|
||||
{"title": "Quantum Computing Principles", "summary": "Quantum entanglement in supercomputers.", "doi": "10.1000/456"}
|
||||
])
|
||||
res = rag.search("How to break LLM hallucination?")
|
||||
print(f"向量化 RAG 检索结果: {res}")
|
||||
Reference in New Issue
Block a user