Agentic AI & RAG Systems
Hands-on builds in agent orchestration and retrieval-augmented generation — a self-correcting LangGraph agent, a Qdrant-backed RAG pipeline with page-level citations, and the same pipeline rebuilt behind a job queue.
A personal repository where I work through agentic AI properly rather than just calling a chat API — building the orchestration, retrieval, and evaluation pieces myself to understand how they fit together.
A self-correcting agent graph. Built with LangGraph, it classifies an incoming query into a typed boolean using structured Pydantic output, routes it down a coding or general branch through a conditional edge, then grades its own answer. If the accuracy score comes back under 80, the graph loops back and refines instead of returning a weak response — a real feedback cycle rather than a single forward pass.
A retrieval pipeline with citations. PDFs are loaded, split into 1000-character chunks with 400 characters of overlap, embedded with text-embedding-3-large, and stored in a Qdrant vector database. At query time a similarity search assembles the context, and the system prompt constrains the model to answer only from what was retrieved — and to point the reader at the exact page number, so answers stay checkable.
The same thing, made asynchronous. Embedding and inference are slow enough to block a request, so the pipeline was rebuilt behind a FastAPI service with an RQ job queue: posting a query enqueues work and returns a job id immediately, a background worker runs the retrieval and generation, and the client polls for the result. Qdrant and the queue both run under Docker Compose.
Repository: github.com/ujjawal2700/GenAI