Kenoscope home
  1. Ask a questionType it or tap an example
  2. Check the answerClick any sentence
  3. Use your documentsCreate a free workspace

Sample library: the GitLab Handbook (CC BY-SA 4.0). Kenoscope is not affiliated with GitLab. Questions are stored to chart gaps; do not enter personal details.

Question answering for your company’s own documents

See what your documents know. And what they don’t.

Ask in plain words. Kenoscope answers only from your documents, shows the exact passage behind every sentence, and tells you when the documents do not have the answer. Try it below on a real team handbook of 53,392 passages.

  • Hosted onMicrosoft Azure
  • Answers byOpenAI models
  • Built withClaude
Scroll to see how an answer is built

Every point of light is one passage from the documents.

3,362 pages were split into 53,392 passages of about 85 words. Passages about the same subject sit close together. The map has 48 topics, and each topic's name comes from the documents' own headings.

Ask a question. The evidence lights up first, then the answer is written.

This is a real answer from this library, replayed at the speed it happened. Each sentence is checked against the passage it cites, then marked supported or not supported.

Recorded on 2026-11-17 with gpt-4.1-mini.

If a sentence says more than its source, it is marked.

Every sentence goes through seven checks in code, then a second reading by the model. In this example the facts are in the passage, but the passage is about one country and the sentence speaks for everyone, so it is marked not supported.

This is a real sentence from our tests. What each check does

What the documents cannot answer is shown on the map too.

When the documents hold no answer, Kenoscope says so and marks where the question landed. Questions that land close together become a named gap, so the owner of the documents knows what to write next.

  • Argentina · Vietnam
  • London · Club
  • Salary · CEO
  • Fy28 · Revenue
  • Siemens · Seat
  • and 14 more

What Kenoscope is for.

Getting an answer you can trust out of a large set of your own documents, without reading them all and without relying on a guess.

It does not search the web, and it does not answer from general knowledge. If your documents do not say it, Kenoscope tells you so.

Try it on the sample handbookUse your own documents

  • Staff policies. Someone asks “How many days of parental leave do I get?” and gets the answer in the policy’s own words, with the page it is on.
  • Customer support. An agent answers from the product manual and can quote the exact passage to the customer.
  • Compliance and legal. See exactly what a policy says, and what it does not cover, before you rely on it.
  • New colleagues. Find the rule without knowing which of a thousand documents holds it.
  • Whoever owns the documents. See which questions people ask that nothing answers, so you know what to write next.

110 of 115 test questions pass.

Every release is run against 115 hand-written questions on this same library: ones it should answer, ones it should refuse, and ones that try to trick it. A release that fails a safety check, or scores more than 3 questions below our best release, does not go live.

See every question and result

  • answered from the right document 78 of 82
  • correctly refused 22 of 23
  • trick questions it did not fall for 10 of 10
  • failed, each listed with the reason 5
sentences marked supported that failed a second, independent check
0 of 183
time to the first words (95 of 100 answers are faster)
1,438 ms
time to the complete, checked answer (95 of 100 are faster)
4,807 ms

Use it from your own software with one request.

Send a question. You get back the answer, the passage behind every sentence, and a verdict for each one. You can also stream the answer word by word.

Read the API guideOpenAPI description

curl https://kenoscope.com/v1/ask \
  -H "Authorization: Bearer $KENOSCOPE_KEY" \
  -H "Content-Type: application/json" \
  -d '{"question": "Can I expense a co-working space?"}'
{
  "status": "answered",
  "claims": [{
    "text": "…",
    "verdict": "supported",
    "citation": {"document": "…", "section": "…",
                 "quote": "…", "url": "…"}
  }],
  "counts": {"claims": 2, "supported": 2},
  "timings_ms": {"search": 41, "first": 44, "total": 46}
}

Your documents stay private to your organisation.

These rules are enforced by the database and by code. None of them depends on a model's judgement.

  • One organisation cannot read another's rows. Every row carries its organisation, and Postgres itself refuses the rest.
  • Keys and passwords are never stored. An API key is shown once and kept only as a hash; passwords only as salted scrypt hashes.
  • A document is data, never an instruction. Text in a question or a passage cannot give the engine orders: 10 of 10 trick questions were stopped in the latest test run.
  • No answer is shown unchecked. If the model, the search or a check cannot be reached, you get an error or a “not supported” mark instead.

Start free. Every plan has every feature.

Cited answers, a verdict for each sentence, the map, the gaps and the API are in all plans. Plans differ only in size.

Free

$0

Try it on a real set of documents.

  • 2,000 passages
  • 200 questions a month
  • 1 library · 3 members · 1 API key

Team

$79 a month

For a team that answers from its own documents every day.

  • 100,000 passages
  • 5,000 questions a month
  • 10 libraries · 25 members · 10 API keys

Scale

$399 a month

For large libraries and products built on the API.

  • 1,000,000 passages
  • 50,000 questions a month
  • 50 libraries · 200 members · 50 API keys

No card is taken on this site. To move to a paid plan, send a request from Settings in your workspace; your plan changes when we approve it.Compare plans

Common questions.

What is Kenoscope for?

Answering questions from your own documents: handbooks, policies, manuals, contracts, guides. You upload them, people ask in plain words, and every sentence of the answer comes with the passage it rests on. When the documents do not hold the answer, it says so instead of guessing.

Which models write and check the answers?

Azure OpenAI (gpt-4.1-mini) writes each answer and does the second reading. An Azure OpenAI embedding model places passages by meaning. The seven code checks use no model at all.

Which files can I upload?

Markdown, Word (.docx), PDF, HTML and plain text, up to 20 MB a file. A 10-page policy is usually 40 to 60 passages.

Is the sample library yours?

No. It is the GitLab Handbook by GitLab B.V., used under CC BY-SA 4.0 because it is a large, real, openly licensed body of team documentation. Kenoscope is not affiliated with or endorsed by GitLab.

Do you keep the questions I ask?

Yes: each question is stored with its outcome, to chart gaps and to count usage against a plan. On the public sample library, do not enter personal details.

Can I check the accuracy figures myself?

Yes. The accuracy page lists every test question, what was expected, and what happened, including the ones that fail.

Does it work on a phone, or without WebGL?

Yes. The map, the question box and the answer sheet are built for touch. A device with no WebGL, or with reduced motion switched on, gets the same map in 2D with the same answers, evidence and gaps.

Try it with your own documents.

Upload Markdown, Word, PDF, HTML or text. The free plan holds 2,000 passages.

Create a workspaceAsk the sample librarySee plans