A public RAG system built to show the parts most demos skip.
github.com/zainasher42/fastapi-docs-ragMost public RAG demos prove that a framework works on a toy dataset. They skip the parts that decide whether a real system holds up.
Vector search and PostgreSQL full-text search each fail in a different direction: embeddings miss exact identifiers and API symbol names, full-text misses paraphrase. RRF merges both ranked lists without needing a tuned weight between them, which matters because the right weight differs per query and there is no principled way to pick one globally.
Chunks split on heading boundaries and carry their section path as a prefix, so every chunk is self-describing. A chunk that reads 'Dependencies > Sub-dependencies > Caching' before its content retrieves correctly for queries about caching in a way that a bare paragraph of prose does not.
Prompts are XML-structured with the stable content first and the volatile content last, so the cacheable prefix stays byte-identical across requests. Structuring for the cache boundary rather than for readability is what brings ingest cost down to roughly five cents for the full corpus.
The golden-set suite covers 18 question/answer pairs and currently passes 94%. The failing cases stay in the report rather than being tuned away or removed from the set. A suite that always passes measures nothing; the value is in watching which cases break when retrieval changes.