Agentic RAG Study

Prototype implementation following ADR Agentic RAG Architecture (LangGraph × Chainlit × Azure OpenAI GPT-5).

Getting Started

# 1. Create a virtual env & install deps
python -m venv .venv
source .venv/bin/activate
pip install -e .

# 2. Run a single-shot demo
python -m src.main "Explain LangGraph in one paragraph"

Architecture

See adr.md for the full decision record. The current codebase only implements a stubbed end-to-end flow so that the graph executes without external services. Replace stub modules under src/ with real integrations as needed.

src/
├─ graphs/          # LangGraph state machine
├─ agents/          # Planner, Synthesizer, Critic
├─ retrievers/      # Web & Vector search
├─ graders/         # Evidence grading
├─ hitl/            # Human-in-the-loop gate
├─ memory/          # Checkpointer & summarizer (TODO)
├─ services/        # LLM and utilities (TODO)
└─ ui/              # Chainlit app (TODO)

Roadmap

1. Replace web/vector retrievers with Azure Bing & AI Search calls.
2. Implement proper evidence grading and confidence scoring.
3. Wire up Chainlit front-end with HITL actions.
4. Add Redis/SQLite checkpointer and conversation memory.
5. Containerize with Docker & add CI.