poppyAI

PoppyAI — Agentic Document Intelligence Platform

Upload your documents. Ask questions. Get answers with citations back to the exact source.

PoppyAI is a production-shaped RAG system, not a notebook demo. A LangGraph agent decides how to answer each question — search your documents, search the web, or break a complex question into sub-questions — then grades what it retrieved and rewrites its own query if the results are weak.

It's built as a polyglot system on purpose: a NestJS gateway handles uploads and API traffic, a Python/FastAPI worker does the AI work, and RabbitMQ sits between them so a 200-page PDF doesn't block a web request.

What makes it different from a "chat with your PDF" demo:

• Every agent step is traced with Langfuse
• Answer quality is scored with Ragas (faithfulness, relevancy, context precision) against a golden dataset — so changes can be measured, not guessed
• Two-level caching (exact + semantic) to keep token costs down
• Custom MCP server exposing tools to the agent
• Kubernetes manifests with proper stateful/stateless separation

#How to start the Project

cd infra/compose docker compose up -d

check: docker compose ps

RabbitMQ UI -> http://localhost:15672 (guest / guest)

Qdrant UI -> http://localhost:6333/dashboard

To Run ingestion-consumer and query-api

Terminal A — ingestion consumer

cd apps/worker && source .venv/bin/activate python -m worker.consumer

Terminal B — query API

cd apps/worker && source .venv/bin/activate uvicorn worker.api:app --port 8000 --reload

#Final Step

Start the Nest JS server cd apps/gateway npm run dev

1) Upload (async ingestion shuru)

curl -F "file=@sample.pdf" http://localhost:3000/documents

-> {"jobId":"a1b2...","documentId":"c3d4...","status":"processing"}

2) Status check (thodi der baad)

curl http://localhost:3000/jobs/a1b2...

-> { ..., "status":"done" } (ya "processing" / "failed")

3) Query

curl -X POST http://localhost:3000/query
-H "Content-Type: application/json"
-d '{"question":"What is the total invoice amount?"}'