An Agentic RAG Web UI system built on the ReAct paradigm with PageIndex, visualizing the document retrieval and reasoning process
dependencies:
pip install -r requirements.txt
start the service
python main.py
visit http://127.0.0.1:8000/
- 🚀 FastAPI-Powered: The service is built on FastAPI, easily integrable into your own applications
- 🗺️ TOC-Driven Navigation: Abandon blind vector matching; the Agent reads the global outline and accurately locates target chapters based on logical relevance
- 🕵️♂️ ReAct Thinking Engine: Empowers the model with autonomous decision-making capabilities, loading content on demand to significantly reduce Token consumption
- ⚖️ Dynamic Evaluation & Reflection: Automatically assesses the validity of extracted text, dynamically builds a high-quality local knowledge base, and filters out irrelevant noise
- 🔌 Plug & Play Simplicity: Natively adapts to the
_structure.jsonformat output by PageIndex, lightweight and ready-to-use, compatible with all OpenAI-compatible APIs
Note: The core PageIndex code in this project has been modified to adapt to OpenAI-compatible interfaces, configured in the .env file (the original project only supports the ChatGPT API)
.env
OPENAI_API_KEY=your_openai_api_key
OPENAI_MODEL=model_name
OPENAI_BASE_URL=base_url
## 🛠️Core Architecture
User Query
│
▼
[Load Document Outline (TOC)] ──────┐
│ │
▼ ▼
Thought-Agent ◄──────────────── [Current Knowledge Base]
│ (Autonomously decide next action) ▲
▼ │
Call Tool (get_texts) │
│ │
▼ │
Judge-Agent ───────────────────────────────┘
(Evaluate result validity and summarize experience)
│
▼ (Triggered when sufficient information is collected)
Call Tool (get_answer)
│
▼
Generate Final Answer
- Support multi-document Q&A capabilities
- Package as MCP interface
- Integrate database storage capabilities
- Connect to Ollama/VLLM to achieve full local deployment


