RAG Source Receipts: offline content and processing fingerprints for Crawl4AI outputs (early MVP)

[Ishaan1342]

I’ve open-sourced RAG Source Receipts, a small MIT-licensed Python sidecar for Crawl4AI-shaped outputs:

https://github.com/Ishaan1342/rag-source-receipts

The problem is deciding whether a previously ingested page needs re-embedding without confusing a new observation time with a new network fetch.

The prototype records SHA-256 fingerprints of the selected text and processing recipe, requested/final URLs, the selected Markdown variant, and separate observation/fetch/cache timestamps. The comparison reports a content change, processing change, or metadata-only redirect change. A timestamp-only change does not trigger re-embedding.

git clone https://github.com/Ishaan1342/rag-source-receipts.git
cd rag-source-receipts
python -m source_receipts demo
python -m source_receipts crawl-result examples/crawl-result.json

Python 3.11+, standard library only. The README includes an object/dictionary adapter example. It rejects failed crawls and missing selected Markdown variants; unknown fetch times stay unknown rather than being inferred from cached_at.

The adapter is tested with fixtures, not a full live Crawl4AI browser run. It does not crawl, embed or sign receipts; hashes do not prove source authenticity. This is an independent project, not an upstream-endorsed integration. GitHub CI runs on Python 3.11 and 3.13.

This explores a portable sidecar version of the use case in #2307. I’d welcome concrete examples where the timestamp semantics or processing fingerprint would be insufficient for your ingestion pipeline, especially cache fallback cases.