Titanic Historical RAG

The killer feature: contradictions, not consensus

Most RAG systems try to return one answer. The witnesses in the Titanic inquiries gave conflicting accounts on almost everything — speed, lifeboat counts, ice warnings, who gave which order. This tool embraces that. Ask “How many people were in Ismay’s lifeboat?” and you’ll see Ismay’s “about 45” set next to Officer Lowe’s “only twelve”, with a confidence-scored explanation of why the statements conflict. And because both the US and British inquiries are indexed, it also catches the 20+ witnesses who testified in both proceedings — and contradicted themselves.

Architecture

What gets indexed

• US Senate Inquiry (1912) — 1,173 pages, fully ingested. 70 witnesses, fully attributed.
• British Wreck Commissioner’s Inquiry — 2,253 pages, fully ingested. 97 witnesses, fully attributed (its transcript format needed a separate boundary parser).

Every citation maps back to the original printed inquiry page — after correcting for PDF pagination drift, which runs 4–10 pages off in the US materials and up to 2.5× off in the British ones.

Measured retrieval

A 30-query gold-set eval keeps the retrieval honest: Hit Rate@5 of 70% (21/30 queries surface a relevant witness in the top five), MRR 0.46 (first relevant hit averages rank 2), and p50 search latency of 276 ms. The eval report documents all 9 misses with an analysis of why each failed rather than burying them. 112 passing tests cover the pipeline.

API shape

POST /search/contradictions returns a list of pairwise contradictions with witness names, the two conflicting claims, a 0–1 confidence score, a one-sentence explanation, and the source chunks behind each claim.

{
  "witness_a": "Joseph Bruce Ismay",
  "witness_b": "Harold Godfrey Lowe",
  "claim_a": "Ismay's lifeboat had approximately forty-five people",
  "claim_b": "Ismay's lifeboat (Boat C) had only twelve people",
  "confidence": 0.95,
  "explanation": "The two witnesses provide directly conflicting specific counts: forty-five versus twelve, which cannot both be true."
}

Roadmap

• Done — Killer feature shipped end-to-end and deployed at titanic.higuera.io (search → LLM verdicts → side-by-side UI behind a custom domain).
• Done — Full British Inquiry ingested (2,253 pages, 97 witnesses) with cross-inquiry self-contradiction detection; DynamoDB-backed verdict cache and rate limiting live in production.
• Next — Reranking ablation with a BGE cross-encoder; contradiction-detection eval against manually labeled pairs.
• Soon — Ship/role filters on existing metadata; grow the gold eval set from 30 to ~200 queries.