Local RAG without magic: sources, timings, request_id, and refusal of generation

▤ Summary

This article outlines a local RAG assistant built with FastAPI, Ollama, and an in-memory vector store. It explains the full backend pipeline from document loading and chunking to embeddings, retrieval, prompt assembly, and answer generation. It also highlights operational features such as request_id tracing, source citations, timing breakdowns, exact-term guards, and manual index rebuilds. The article emphasizes production-readiness concerns like stale index handling, retrieval thresholds, and hybrid semantic plus keyword search.

▦ Linked Companies