Every answer, marked at the source.

Ask the Express 5 documentation a question. DocuMind answers with the exact sentence, highlights it on the original page, and says "not in the docs" when it can't.

No account and no API key. Uploaded files are never written to disk.

Try it liveSee the measurements

guide / routing

One question, four checks.

  1. 01
    SplitPages are cut at headings. Each passage keeps its page title and heading.
  2. 02
    RetrieveKeyword search and a semantic index vote. A reranker orders the best five.
  3. 03
    CheckThe best sentence must cover enough of the question, or there is no answer.
  4. 04
    MarkThe sentence is copied, cited, and highlighted on the original page.

Coverage of the question by the best sentence

Small enough to read. Measured before it was trusted.

Each stage was kept or dropped because a number said so.

Chunk

Pages are split at headings. Each passage keeps its page title and heading, so a bare paragraph still has context.

198
passages from 16 pages

Retrieve

Keyword search and a local semantic index each vote. A reranker orders the top five.

71.4%
right passage ranked first

Answer or refuse

Only from those passages. Unknown words or weak coverage mean a refusal, not a guess.

15 / 15
uncovered questions refused

The numbers, including the one I'd rather hide.

Fifty questions written for this corpus: 35 the docs answer, 15 they don't.

Right passage ranked first

retrieval recall@1 · 35 answerable questions

Semantic vectors
40.0%
LSA only
54.3%
Keyword + semantic
57.1%
Keyword (BM25)
65.7%
Keyword + reranker
68.6%
Hybrid + reranker
71.4%

Answering, held-out half

extractive mode · settings chosen on the other half

100%

of uncovered questions refused. Zero false answers.

29%

of answerable questions answered correctly. Safe and conservative, not smart.

Pending

Model-written answers and prompt-injection results need an API key. No number is claimed until they run.

Errata: what this does not claim.

Read these before trusting any number above.

  1. 1Small benchmark. The questions were written by the author for one corpus. With 35 answerable questions, one question is 2.9 percentage points.
  2. 2Paraphrases stay hard. Two questions are missed by every local method because they share almost no words with the answer. That is the case for a neural embedding model.
  3. 3Thresholds are tuned on Express. On your own documents the refusals may be stricter or looser than ideal. I haven't measured it.
  4. 4This site is extractive. Until a model key is set on the server, it copies the best sentence. It doesn't write answers, so it can't invent one.

Ask it something it can't answer.

Open the app