Developer Builds Hybrid RAG API Using Semantic Search, Reranking, and Caching

A software developer has detailed the architecture behind a production-ready PDF question-answering API built with FastAPI, Qdrant, PostgreSQL, Redis, and LiteLLM. The system combines dense semantic embeddings with sparse keyword vectors to handle both meaning-based and exact-term queries, addressing a key limitation of purely semantic retrieval. When a user submits a question, the API retrieves 20 candidate chunks via reciprocal rank fusion before a cross-encoder reranker narrows them to the top five for the LLM prompt. PDF uploads are processed asynchronously, with text extraction and vectorisation handled in background tasks to avoid blocking API responses. Redis caching and source metadata preservation are also built into the pipeline, making the LLM call the final and least complex step in the workflow.