Author(s): Armin Norouzi, Ph.D

Originally published on Towards AI.

Production RAG API with FastAPI, pgvector, and Claude

Most RAG tutorials show you how to call an embedding API and do a similarity search. That is the easy 20%. The hard 80% is the architecture around it: how you chunk documents, how you prevent retrieved chunks from being redundant, how you expose this as a reliable HTTP API, and how you reason about the latency budget at each stage.

The article walks through building a production-ready RAG system end-to-end: an architecture overview of a sequential pipeline (chunking, retrieval, MMR reranking, and LLM prompting), why chunking strategy is critical (token-window size and overlap tradeoffs), and a complete pure-Python implementation using TF-IDF and cosine similarity with a greedy MMR diversity strategy. It then discusses how to profile latency and optimize the right stages (generation dominates), maps each simulated component to production equivalents (sentence-transformer embeddings and a pgvector backend), and explains practical tuning recommendations such as starting with chunk_size≈150/overlap≈30, using λ≈0.5 for MMR, retrieving 3× candidates before reranking, monitoring latency percentiles, and applying metadata filters before ANN search.

Read the full blog for free on Medium.

Join thousands of data leaders on the AI newsletter. Join over 80,000 subscribers and keep up to date with the latest developments in AI. From research to projects and ideas. If you are building an AI startup, an AI-related product, or a service, we invite you to consider becoming a sponsor.

Published via Towards AI