chain-rag — Cited Retrieval over Blockchain Protocol Docs
A second, independent vertical on the same architecture — cited answers over Bitcoin, Ethereum, Solana, Monero, Polygon, and Cardano protocol docs — plus a real router-calibration finding the reliability gates exist to catch.
- 22 docs / ~1,055 chunks
- real corpus across 6 chains measured
- 8/8
- eval checks (6 per-chain + 1 cross-chain + 1 abstain) measured
- 3 PASS / 0 FAIL
- reliability gates (VecStamp, ChunkLedger, and more) measured one gate N/A for this corpus
Context
A design that only works on one corpus is a coincidence, not an architecture. chain-rag rebuilds the Consilium approach on a completely different domain — blockchain protocol documentation — to see whether it holds up. It does, and it surfaced a real calibration issue in the process.
What I built
- A second, independent vertical — cited retrieval across six chains’ protocol docs (Bitcoin, Ethereum, Solana, Monero, Polygon, Cardano).
- The same reliability gates run against a new corpus, one of which flagged a real router-calibration finding: dense-embedder anchor thresholds need per-corpus tuning.
Why it matters
Eight of eight eval checks pass — six per-chain, one cross-chain, one abstention — on a real 22-document corpus. The blockchain angle isn’t the point. The point is the architecture held up on a new corpus, and the gates caught a real calibration bug instead of just passing again.