Citation-grounded RAG
Answers you can verify, line by line.
A retrieval-augmented answer engine where every claim is grounded in a specific source passage — with inline, checkable citations and an abstain path when the corpus doesn't support an answer. Built for high-stakes text, where a confident wrong answer is worse than no answer.
The demo corpus is the U.S. founding documents — the Constitution, the Bill of Rights, and the Federalist Papers (public domain) — so every answer is externally checkable.
The First Amendment bars Congress from establishing a religion or prohibiting the free exercise thereof.1
Congress shall make no law respecting an establishment of religion, or prohibiting the free exercise thereof; or abridging the freedom of speech, or of the press; or the right of the people peaceably to assemble…
- 1.00
- faithfulness · Claude judge
- 3/3
- abstain · out-of-corpus
- 0.909
- recall@5 · hybrid RRF
- 0.809
- MRR · hybrid RRF
How it works
Two readings, fused — then grounded, or withheld.
Retrieval finds the passages; generation is constrained to cite them. The whole path is scored by an eval harness that gates CI on retrieval recall and answer faithfulness.
01Hybrid retrieval
BM25 lexical and dense embeddings, fused with Reciprocal Rank Fusion — exact terms, names, and §-numbers from one; paraphrase and concept matches from the other. Neither half alone is enough.
02Native, verifiable citations
Sources go to Claude as document blocks with citations enabled, so every cited span is a real substring of a real passage — not a model-invented [1]. Open any marker to read it in context.
03Abstains, doesn't guess
When retrieval surfaces no supporting evidence, the engine declines instead of fabricating. An out-of-corpus question is met with silence, not confidence — measured at 3/3 abstain accuracy.
Hybrid retrieval
Ask it something checkable.
Query the founding documents and watch every claim attach to the exact passage that supports it — or watch it decline when the sources don't.