facti API

Most search engines count links. facti checks the evidence.

Every result carries a reliability rating judged from the page itself — citations, attribution to named sources, dates, corrections, and corroboration across independent domains. A program can use it to decide how much weight to give a source before quoting it.

Request

GET https://facti.co.uk/api/v1/search?q=your+query&k=10&rel=good
Authorization: Bearer YOUR_KEY

parameter	default	meaning
q	required	the query
k	10	results wanted, 1–50
rel	good	high — only well-evidenced pages good — exclude poorly evidenced ones any — everything

Response

{
  "query": "uk inflation rate",
  "standard": "good",
  "count": 2,
  "withheld": 3,
  "results": [
    {
      "title": "Consumer price inflation, UK",
      "url": "https://www.ons.gov.uk/economy/inflationandpriceindices",
      "snippet": "Price indices, percentage changes and weights for the...",
      "reliability": "high"
    }
  ]
}

withheld is how many results the evidence bar removed. count: 0 with withheld: 6 means we found things but none met the bar — retry with rel=any before concluding nothing exists. That distinction is deliberate; use it rather than guessing.

count: 0 with withheld: 0 means we hold nothing about your question, and rel=any will not help. This is a deliberate answer, not a failure — please do not retry it as though it were an error.

We would rather say nothing than guess. A search engine that always returns something returns confident nonsense for anything outside what it holds, and a well-evidenced page that is not about your question is the worst kind of wrong answer: a model will cite it. So a result set has to be about the question before we will show it. If you need breadth we cannot give, send the query to another source when count is zero.

What reliability means

rating	how to treat it
high	cite it — well evidenced, or a primary source
medium	usable but unproven. Most ordinary pages sit here, including good ones that simply have nothing to cite. Not a warning.
low	do not cite without corroboration

It is not a claim that the content is true. We assess evidence of factual care, not truth. Please do not present it to your users as “verified” — that is an overclaim, and it would rightly cost us both trust the first time a well-evidenced page turned out to be wrong.

Where facti is strong — and where it is not

The index is curated depth, not web breadth: sources chosen for reliability rather than whatever ranks. It is strong on standards and RFCs, official software documentation, UK public data and legislation, science and health guidance, economics, and law.

It is not a general web search. Current news, consumer topics, product information, anything local — use something else, or use facti as the authoritative-source lane alongside a general engine. Saying so plainly is worth more to you than a claim we would fail to keep.

Calling it from a model directly

An MCP server ships with the client, so your model can call facti itself rather than you wiring it up:

{
  "mcpServers": {
    "facti": {
      "command": "python",
      "args": ["-m", "arx52search.mcp_server"],
      "env": { "FACTI_API_KEY": "YOUR_KEY" }
    }
  }
}

Errors

status	meaning
401	missing or unrecognised key
403	key revoked
429	monthly quota exceeded
503	our fault — pause and retry once, please do not hammer

Getting a key

Keys are issued by hand while facti is in private preview — hello@facti.co.uk. The free tier is 10,000 queries a month, set deliberately high enough to build with rather than merely evaluate.