rag-tut

A small modular RAG service: ingest PDFs into Qdrant, retrieve/answer over them with Ollama, exposed via a FastAPI API and a dark-mode Streamlit chat UI. LangChain runs are traceable through LangSmith.

Layout

app/
  config.py          # central settings from .env
  observability.py   # LangSmith tracing setup
  main.py            # FastAPI app + startup
  api/
    routes.py        # /ingest and /extract endpoints
    schemas.py       # request/response models
  services/
    rag.py           # ingestion + retrieval + answer generation
streamlit_app.py     # dark-mode chat interface
.streamlit/config.toml
ingest.py / extract.py / llm-judge.py   # original CLI scripts (still usable)

Run

1. Start Qdrant: docker compose up -d
2. Start the API: uv run uvicorn app.main:app --reload
   • Docs at http://localhost:8000/docs
3. Start the chat UI: uv run streamlit run streamlit_app.py

Endpoints

• POST /ingest — multipart upload of a PDF (file, optional clean flag).
• POST /extract — JSON { "query": str, "k": int, "generate": bool }. With generate: true it returns a grounded answer plus source chunks.

Observability (LangSmith)

Add to .env to capture LangChain traces:

LANGSMITH_TRACING=true
LANGSMITH_ENDPOINT=https://api.smith.langchain.com
LANGSMITH_API_KEY=<your LangSmith key>
LANGSMITH_PROJECT=rag-tut

Startup logs whether tracing is active.