How is this different from a vector database?

A vector database finds text that is similar. We check whether a specific passage supports a specific claim, and return the printed page number so the attribution can be checked by a human.

What if the corpus does not cover my domain?

You get explicit misses, free of charge, and can trigger a harvest for that topic. The corpus is shared, so once a subject is fetched it stays for everyone.

Can I get the full text?

A short quotation under fair use, plus a link to the original. For more than half the corpus that link opens the full text, because those are open-access publications.

Ground a sentence in a source that actually says it

Send the claim your model just wrote. Get a real publication, the printed page the statement appears on, and the passage itself — or an explicit miss when nothing in the corpus supports it.

Get a test key Documentation

Why retrieval is not enough

Vector search returns documents that are about the same topic. That is a different question from whether a document supports this particular sentence — and the gap between the two is where fabricated attribution comes from. A paper on screen time in twelve-year-olds is topically perfect for a claim about preschoolers and still cannot support it. Grounding means answering the second question, and answering it with a page a reader can turn to.

What happens to a claim

1
Full-text search over the corpus
Not metadata — the text of the documents themselves, split by printed page.
2
The passage is judged against the claim
Not against the topic. Same phenomenon in a different population scores lower and says so.
3
You get the page, not the PDF sheet
The number printed on the page — the one a reader checking the footnote will land on.

POST /api/ext/find-source

Up to 60 claims per request. The hint field carries the topic of the work, which turns the question from “does this support the sentence” into “could someone writing this paper cite it”.

curl -s -X POST https://cytado.com/api/ext/find-source \
  -H "Authorization: Bearer $CYTADO_KEY" \
  -H "Content-Type: application/json" \
  -d '{
  "items": [
    {
      "id": "a1",
      "claim": "Prokrastynacji towarzyszy poczucie winy i dyskomfort psychiczny.",
      "hint": "prokrastynacja u studentow"
    }
  ],
  "langs": [
    "pl"
  ],
  "harvestOnMiss": false
}'

Full reference, every field and error code →

What a miss means

found: false means the corpus has nothing to support this claim — and we return that instead of the nearest match dressed up as an answer. Misses are free. If you want us to go and fetch sources for a subject we do not have yet, that is a separate, deliberate call: harvestOnMiss takes minutes rather than seconds and is metered separately, because building a corpus for a new topic costs two orders of magnitude more than answering from one.

FAQ

How is this different from a vector database?: A vector database finds text that is similar. We check whether a specific passage supports a specific claim, and return the printed page number so the attribution can be checked by a human.
What if the corpus does not cover my domain?: You get explicit misses, free of charge, and can trigger a harvest for that topic. The corpus is shared, so once a subject is fetched it stays for everyone.
Can I get the full text?: A short quotation under fair use, plus a link to the original. For more than half the corpus that link opens the full text, because those are open-access publications.

Point it at your worst hallucination

Take a sentence your model produced with a confident citation attached, and see whether anything real supports it. 300 free checks with a test key.

Create an account

Looking for something else? Check whether a citation is real · A whole reference list · the API overview