RAG Laboratory

Visual workflow editor for a real AI/RAG pipeline. The canvas is a debugger on top of a FastAPI execution engine. Node colors, counters, confidence, latency, retrieved chunks, and LLM text come from actual operations — nothing is simulated in the UI.

Stack

• Frontend: React, TypeScript, Vite, React Flow, Tailwind
• Backend: Python, FastAPI, SQLAlchemy, SSE
• Data: PostgreSQL + pgvector (one database)
• Default model host: local Ollama

Prerequisites

• Docker (for PostgreSQL)
• Python 3.11+
• Node.js 20+
• Ollama running locally

Pull at least one chat model and one embedding model:

ollama pull qwen2.5
ollama pull nomic-embed-text

nomic-embed-text is 768-dimensional, which matches EMBEDDING_DIM in .env.

Run

cp .env.example .env
docker compose up -d
cd backend
python -m venv .venv
# Windows: .venv\Scripts\activate
pip install -r requirements.txt
cd ..
uvicorn app.main:app --reload --app-dir backend --port 8000

In another terminal:

cd frontend
npm install
npm run dev

Open http://localhost:5173.

NVIDIA and GitHub Models are optional. Set NVIDIA_API_KEY or GITHUB_TOKEN in .env (server-side only). Switch the provider on an LLM or Embedding node without changing the graph.

Demo workflow

On first boot the backend seeds a 3-page PDF (energy_efficiency_practices.pdf) and a workflow with:

1. PDF → chunker → embeddings → PostgreSQL/pgvector
2. User query → orchestrator → prompt → LLM → confidence → branch
3. Low confidence → query embedding → vector search → prompt → LLM → confidence → output

Use Run workflow. Click a node for input/output/telemetry. Click an edge for the payload that actually crossed it. Delete the Vector Search node and run again: if confidence is below the configured threshold, the run fails with a retrieval-path error.

Notes

• Query embeddings are not stored as knowledge. Document embeddings are.
• Confidence is a weighted combination of observable signals (completeness, consistency, evidence coverage, answer validation), not an LLM self-score.
• Pause / resume / stop are honored between nodes, not mid-token.

RAG-PLAYGROUND