EduPolicy Agent — Agentic AI Assistant for University Regulations

Due to difficulties in finding out and search for suitable regulations attachments when problem arise, I find this is a very complex due to the disorder and organize of how these documents are stored, therefore I create this agent, to help me search for relevent infomation with some simple prompt, no need to look hopelessly for hours just to find some very short answer.

Features

• ReAct Agent Orchestration – multi-tool reasoning loop powered by LangChain and OpenRouter LLMs.
• RAG over University Regulations – PDF ingestion, chunking and embedding (intfloat/e5-large-v2) stored in Milvus for semantic retrieval.
• NL2SQL Analytics – Natural language questions translated into SQL against student_records.db for academic performance queries.
• Live Web Search – Tavily integration to pull in Ministry and education guidance when internal knowledge is insufficient.
• Explainable Responses – reasoning trace and tool observations surfaced in the UI for auditability.
• Streamlit UI – lightweight chat interface for demonstrations or can be replaced by OpenWebUI.
• Dockerised Deployment – Dockerfile + docker-compose for reproducible setup including Milvus standalone.

Project Structure

edupolicy-agent/
├── app/
│   ├── main.py                # FastAPI entrypoint
│   ├── config.py              # Pydantic settings
│   ├── schemas.py             # Pydantic models
│   ├── utils.py               # Utilities & session memory
│   ├── agents/
│   │   ├── controller.py      # ReAct agent orchestrator
│   │   └── tools/             # Tool implementations
│   └── db/                    # Milvus & SQLite clients
├── data/
│   ├── all_regulations_files.pdf        # 
│   ├── student_records.db     
│   └── embeddings/            # Placeholder for exported vectors
├── ui/
│   ├── app_ui.py              # Streamlit UI
│   └── openwebui_config.json  # Configuration
├── requirements.txt
├── Dockerfile
├── docker-compose.yml
└── README.md

Quick Start

1. Clone repository & install dependencies

   python -m venv .venv
   .venv\Scripts\Activate.ps1 (Command Prompt) or source .venv/bin/activate (Bash)
   pip install --upgrade pip
   pip install -r requirements.txt

2. Configure environment

   export OPENROUTER_API_KEY="your-openrouter-key"
   export TAVILY_API_KEY="optional-tavily-key"
   # Optional overrides
   export OPENROUTER_BASE_URL="https://openrouter.ai/api/v1"
   export MILVUS_URI="http://localhost:19530"

3. Start Milvus – run via docker compose or connect to an existing Milvus instance.

   docker compose up milvus -d

4. Run FastAPI server

   uvicorn app.main:app --reload

5. Launch the Streamlit UI (optional)

   streamlit run ui/app_ui.py

   The UI communicates with the FastAPI backend at http://localhost:8000.

Embedding

All the pdf files will be embedded before put into Milvus database, BAAI/bge-m3 embedding model is chosen because of its ability in embedding multilingual, therefore it very suitable to use with Vietnamese. The data/Embedding.ipynb notebook shows how to embed your PDFs into Milvus Lite using the BAAI/bge-m3 model.

Agent Behaviour

• The system prompt enforces tool-aware behaviour, instructing the model to reason in Vietnamese and clearly state when tools are used.
• Toolset: rag_tool, sql_tool, web_tool, summarizer.
• Conversation history is persisted as JSON (data/session_memory.json) keyed by session_id allowing stateless API deployments with lightweight persistence.

Data & Customisation

• Place an institution-specific related-to-law PDF at data/ folder. The repository does not ship a sample for licensing reasons. The ingestion pipeline runs automatically on boot when the Milvus collection is empty and the file is available.
• Fine-tune chunking or retrieval depth via config.py.

Docker Deployment

A production ready stack can be launched with:

docker compose up --build

Services started:

• edupolicy-backend – FastAPI + agent, exposed on 8000.
• milvus-standalone – Vector database with persistent volume.
• edupolicy-ui – Streamlit UI exposed on 8501.

Provide environment variables via .env for secrets. Mount real PDFs/databases using volumes if required.

API Reference

• POST /chat – body {"session_id": "...", "message": "..."}. Returns agent answer, reasoning and tool logs.
• POST /rag/query – semantic search over regulations, returns concatenated context and snippet list.
• POST /sql/query – direct access to SQL tool (useful for testing).
• POST /web/query – execute Tavily search.
• GET /health – health probe.

Testing the Agent

Example question flow once the backend is running:

curl -X POST http://localhost:8000/chat \
  -H "Content-Type: application/json" \
  -d '{"session_id": "demo", "message": "Điều kiện bị cảnh báo học vụ là gì?"}'

The agent will fetch relevant regulation snippets, optionally summarise and respond. Inspect the reasoning trace to verify tool usage.