pranavparuchur1/docintel-rag ? reverse-engineered prompt

Build me an end to end app that ingests SEC filings, breaks them into chunks, turns them into embeddings, and stores them in a versioned pgvector index so I can keep serving one index while building the next.

I want a grounded question answering flow that can look up the right filing, answer with citations, and say it does not know when the evidence is weak instead of making things up. Please make the ingestion idempotent, use content hashes so unchanged chunks are not re embedded, and make the whole thing run locally with Docker Compose and no paid API keys required.

Also add a way to measure retrieval quality against a golden question set, with recall and refusal metrics, and include a simple way to run the pipeline from the command line and expose the app through an API. If you need to check current library docs online, go ahead.