AI Text Summarization & Q&A

Two calls over a block of text: summarise it, or ask it a question. Both answer with structured JSON, and both are constrained to the document you send — the model is instructed to work from that text alone and to say so when it cannot.

POST /api/v1/text-ai
X-Api-Key: ksty_live_...
Content-Type: application/json

{"text": "<the document>", "length": "short"}

{
  "title": "Q3 revenue review",
  "summary": "Revenue fell 12% to $4.1M, driven by churn in the SMB tier...",
  "key_points": [
    "Revenue $4.1M, down 12% quarter on quarter",
    "SMB churn rose to 4.8% monthly",
    "Board approved a hiring freeze through Q1"
  ],
  "usage": { "input_tokens": 1840, "output_tokens": 310 }
}

Grounding — what "from the document alone" buys you

Both endpoints run under instructions to use only the text you send: no background knowledge, no inference, no filling of gaps. That constraint is the point of the API. A general-purpose model asked "what is this company's revenue?" will often answer something — plausibly, fluently, and from the wrong source. Here, text-ai.ask returns "answered": false and tells you what is missing instead, and every answer it does give arrives with the passages it rests on in quotes, so you can verify it without re-reading the document.

This narrows the failure mode; it does not eliminate it. Treat quotes as the check: an answer whose quotes do not appear in your source is one to reject, and that comparison is cheap to automate.

Pricing — why this API is metered per token

Every other API in this catalog charges a flat price because its work is bounded. This one calls a paid language model, where cost scales with how much text goes in and how much comes out, so it is priced the same way:

A token is roughly four characters of English prose. The gateway pre-charges an estimate — your text sized generously, plus the full output ceiling for the length you asked for — and the handler reports what the call actually used, which refunds the difference before the response returns. X-Credits-Charged is the final figure, and usage in the body shows the token counts behind it.

Because the estimate assumes the maximum, a long summary places a larger hold than it usually costs. Expect the refund; do not budget against the hold.

Live keys only, and no per-API free allowance

text-ai has no test sandbox and no monthly free grant. Both follow from the same fact: every call spends real money on a metered model upstream, so a free path to it — whether a sandbox key or a refilling allowance — is not something this platform can leave open. A ksty_test_ key gets 403 sandbox_disabled.

Calls draw on your normal account balance, which includes the free credits granted at signup, so there is nothing to buy before your first call.

Endpoints

Summarise — POST /api/v1/text-ai

Returns a title, a prose summary, and the specific load-bearing facts as key_points. length controls roughly how long the prose runs (short ≈ 60–100 words, medium ≈ 150–250, long ≈ 400–600); it does not change how much of the document is read. Set language to translate the output — the document stays in its own language, the summary comes back in yours.

Ask — POST /api/v1/text-ai.ask

Answers question from the document. Read answered before answer: when it is false, answer explains what the document does not cover and quotes is empty.

Repeated questions against the same document are the pattern this endpoint is built for, and documents over about 4,000 characters are cached upstream for a few minutes so the follow-ups cost noticeably less than the first. Send the same text byte-for-byte to benefit — any edit starts a new document as far as the cache is concerned.

Limits

• 200,000 characters of text per call, about 80 pages of prose. Split longer documents and summarise the summaries; for Q&A, send only the sections that could hold the answer.
• One language per call. A mixed-language document is summarised in whichever dominates unless you set language.
• No URL fetching. Pair with html-parser to turn a page into text first.

When to use it

• Condensing support tickets, transcripts, or long email threads into something scannable.
• Answering a fixed set of questions across a corpus of contracts or reports, with quotes to audit.
• Screening inbound documents before a human reads them — and knowing, via answered, which ones do not contain what you need.