一个 API Key、零向量库、零本地语料
直接从 2.5 亿篇开放学术论文中生成循证医学回答
完全开源地轻量级复现 OpenEvidence 风格的医疗循证问答体验。
不建索引，不存论文，不训模型，一个API就跑起来 — 用查询时编排替代预处理重资产。
💡 为什么是 Lite？ — 除了一个 LLM API，仅依赖 OpenAlex 的 2.5 亿篇开放学术数据。无需向量数据库、无需 Embedding 模型、无需下载处理论文，略过昂贵繁重的数据集处理和embedding，即刻跑通。
- 🪶 无需本地向量数据库 — 无本地语料、无 Embedding 预处理、无索引构建。基于 OpenAlex 2.5 亿篇开放学术文献实时检索，一个 API Key 开箱即用
- 🔍 跨语言三视角检索 — 中文问题自动生成：英文学术检索式 + 自然语言改写 + HyDE 假想摘要，三条路径同时召回，用查询增强替代本地索引
- 🎯 小模型证据门控 — 廉价推理模型并发审读每篇候选，将无关论文挡在大模型上下文之外。看似多一步，实则总 token 更少、延迟更低、回答更准
- 📝 Grounded Answer — 循证结构化回答，引用自动重编号并与源文献对齐，每个结论可溯源到具体论文
- 🧩 Lite 内核，Full 上限 — 统一文档模型 + 六阶段编排 + 可插拔数据源，从零索引平滑升级到本地向量库 + 临床指南的完整 RAG
「一番の近道は遠回りだった。遠回りこそが俺の最短の道だった」 — Gyro Zeppeli, Lesson 5 (《JOJO的奇妙冒险 第七部 飙马野郎》)
小模型全文过滤看似绕了远路，但把噪音挡在大模型上下文之外，节省了回答模型的昂贵context window，反而省时又省钱。
用户提问（中文 / 英文）
│
├─ Stage 0: 安全边界 轻量 LLM 拦截非医学/危险问题
├─ Stage 1: 三视角查询增强 检索式生成 + 英文改写 + HyDE 假想摘要
├─ Stage 2: 开放学术检索 三策略从 2.5 亿篇论文中召回候选
├─ Stage 3: 去重与排序 DOI/标题模糊去重 + RRF 多路融合
├─ Stage 4: 小模型证据门控 廉价模型并发审读，砍掉噪音 ← 真正的捷径
└─ Stage 5: 循证回答生成 强模型精读 + 引用对齐 + 参考文献格式化
git clone https://github.com/yourname/openoe-lite.git
cd openoe-lite
pip install -e .
cp .env_template .env # 编辑：填入 OPENROUTER_API_KEY
python -m openoe # 浏览器打开 http://localhost:8000编辑 openoe/config/pipeline.yaml 自定义模型和各阶段参数。
所有模型通过 OpenRouter 调用，直接替换模型名即可切换。
| 环境变量 | 必填 | 说明 |
|---|---|---|
OPENROUTER_API_KEY |
✅ | 获取 Key |
OPENALEX_EMAIL |
OpenAlex 礼貌池邮箱 | |
OPENOE_DEBUG |
设为 true 记录 LLM 完整请求/响应到 logs/llm/ |
| 方法 | 路径 | 说明 |
|---|---|---|
GET |
/ |
🖥️ Web 前端 |
POST |
/api/query |
RAG 问答 |
WS |
/ws/query |
实时问答 (带进度推送) |
GET |
/api/sources |
检索路线列表 |
GET |
/api/config |
当前配置 (脱敏) |
GET |
/api/traces/{id} |
执行轨迹 |
GET |
/health |
健康检查 |
模块化架构让你可以逐步增强：
- 🗄️ 接入本地论文数据库 (PostgreSQL + pgvector 全文/向量检索)
- 📚 添加临床指南数据库 (chunk 级向量检索 + 段落提取)
- 📥 构建 PubMed XML 数据灌入管线
- 🤖 实现 Agentic RAG (多轮迭代检索)
- 👤 添加用户系统 + 查询历史
- 🪶 No Local Vector DB Needed — No local corpus, no embedding preprocessing, no index building. Real-time retrieval from OpenAlex's 250M+ open academic works — one API key, ready to go
- 🔍 Cross-Lingual 3-View Retrieval — Chinese questions auto-generate: English search expression + natural language rewrite + HyDE hypothetical abstract. Three paths retrieve simultaneously, replacing local indexing with query-time augmentation
- 🎯 Small-Model Evidence Gating — Cheap reasoning models concurrently review every candidate, keeping irrelevant papers out of the strong model's context. One extra step → fewer tokens, lower latency, better answers
- 📝 Grounded Answers — Evidence-based structured responses with auto-reindexed citations aligned to source papers. Every conclusion is traceable
- 🧩 Lite Core, Full Ceiling — Unified document model + 6-stage orchestrator + pluggable sources. Upgrade smoothly from zero-index to full RAG with local vector DB + clinical guidelines
"The shortest route was a detour. It was a detour that was our shortest path." — Gyro Zeppeli, Lesson 5 (JoJo's Bizarre Adventure Part 7: Steel Ball Run)
Small-model filtering looks like a detour — but keeping noise out of the strong model's expensive context window actually saves both time and money.
Question (Chinese / English)
│
├─ Stage 0: Safety Gate lightweight LLM blocks non-medical / unsafe queries
├─ Stage 1: 3-View Enhancement search expression + English rewrite + HyDE hypothetical abstract
├─ Stage 2: Open Academic Search 3 strategies retrieve from 250M+ papers
├─ Stage 3: Dedup & Rank DOI/title fuzzy dedup + RRF multi-path fusion
├─ Stage 4: Evidence Gating cheap models concurrent review ← the real shortcut
└─ Stage 5: Grounded Answer strong model deep-read + citation alignment + reference formatting
git clone https://github.com/yourname/openoe-lite.git
cd openoe-lite
pip install -e .
cp .env_template .env # Edit: add OPENROUTER_API_KEY
python -m openoe # Open http://localhost:8000Edit openoe/config/pipeline.yaml to customize models and stages.
Models are accessed via OpenRouter — swap any model by name.
| Variable | Required | Description |
|---|---|---|
OPENROUTER_API_KEY |
✅ | Get your key |
OPENALEX_EMAIL |
Email for OpenAlex polite pool | |
OPENOE_DEBUG |
true to log LLM requests to logs/llm/ |
| Method | Path | Description |
|---|---|---|
GET |
/ |
🖥️ Web frontend |
POST |
/api/query |
RAG query |
WS |
/ws/query |
Real-time query with progress |
GET |
/api/sources |
Active retrieval routes |
GET |
/api/config |
Current config (redacted) |
GET |
/api/traces/{id} |
Execution trace |
GET |
/health |
Health check |
openoe/
├── pipeline/ # 6 RAG stages (one file each)
│ ├── safety.py # Stage 0: Safety classification
│ ├── enhance.py # Stage 1: HyDE + translate + keywords
│ ├── retrieval.py # Stage 2: Multi-source concurrent retrieval
│ ├── merge.py # Stage 3: RRF fusion + dedup
│ ├── filter.py # Stage 4: LLM relevance filter
│ ├── answer.py # Stage 5: Answer + citation formatting
│ └── orchestrator.py# Wires stages 0→5 with progress callbacks
├── sources/
│ └── openalex.py # OpenAlex 3-way search
├── llm/
│ ├── client.py # Async LLM client (retry, reasoning control)
│ └── prompts.py # YAML prompt loading
├── core/ # Config, models, tracing
├── api/app.py # FastAPI + WebSocket + frontend serving
├── static/index.html # Frontend (single-file, no build step)
├── config/ # pipeline.yaml + prompts.yaml
└── utils/ # Text cleaning + dedup
The modular architecture lets you progressively enhance:
- 🗄️ Add a local paper database (PostgreSQL + pgvector for full-text + vector search)
- 📚 Add clinical guideline retrieval with chunk-level vector search
- 📥 Build a PubMed XML ingestion pipeline
- 🤖 Implement Agentic RAG (multi-round iterative retrieval)
- 👤 Add user system + query history
MIT License · Made for learning and sharing