POST /api/answer searches the web, reads the pages it finds, and answers by
quoting them.
Nothing is generated. Every sentence in answer is a span taken verbatim from
the source its marker points at, so a citation cannot point at something the
page never said. The quoted text and the citation are the same object.
Request
string
required
The question, up to 512 characters.
q is accepted as an alias.string
default:"instant"
Search depth behind the answer:
instant, deep or deep_reasoning. A
deeper tier reads more pages, so there is more to quote from. See
tiers.integer
default:"10"
How many ranked results the answer may quote from. Clamped to 50.
integer
default:"5"
How many quotes the answer carries. Clamped to 12. Past a dozen it stops
reading as an answer and starts reading as the search results again.
string
default:"json"
markdown or text also fill rendered with the answer and its sources as
one document, ready to paste into a note or a prompt.string
Restrict to one domain, so the answer can only quote pages from it.
string
day, week, month or year. Excludes anything older before the answer
is built.string
string
default:"general"
general, news, science, images, videos or it.How the answer is built
Four steps, and each one is visible in the response. Search. The tier’s budget decides how wide recall goes and how many pages get read. Results arrive ranked, with matching passages already scored. Select. Passages are scored on their own match weighted by how well search ranked the page they came from, so a strong quote on a weak page does not outrank a strong quote on the page search put first. Spread. At most two passages come from any one document, so an answer is corroborated rather than one page quoted four times. A question only one document answers still gets a full answer: the cap lifts when nothing else has anything to say. Assemble. Quotes are joined in score order, each with a[n] marker, one
paragraph per statement. Sources nothing quoted are dropped and the rest are
renumbered, so the list is a citation list rather than the search results
again, and every marker resolves.
considered reports how many passages were available before selection, which
is the number to tune passages against. A low passages against a high
considered means you are leaving support on the table.Reading the citations
sources[].n is the number used in answer. Markers are contiguous from 1, so
[3] always exists when it appears.
passages[].source points at the same number, and passages[].text is the
exact span that was quoted. That pairing is what makes an answer checkable:
take any marker, find its passage, and search the source for that text.
Ordering is by score rather than by source, so the best supported statement
comes first. The markers carry provenance, so reading order does not have to
match source order.
What a page with no body contributes
A result Truscan could not read still appears as a source, and its engine snippet can be quoted. Snippets score well below any extracted passage, so they are only used when nothing better was read. A page that failed to fetch does not appear inpages_read, and neither does
a client-rendered shell that yielded almost nothing.
Cost
An answer is one charge. The search behind it and the pages it reads are included in that price, so the same question at a deeper tier reads further and costs the same. That is deliberate. You asked one question. How wide the search goes and how many pages get read to answer it is our retrieval strategy, not something you ordered, and billing it back would make a better answer cost you more without your having asked for one.usage reports the work that price covered, not additional charges:
search_requests and pages_read are 0 when the search behind the answer
was served from cache, because nothing was retrieved. answers is still 1:
the answer was still assembled.
Units rather than dollars on purpose. Multiply
answers by the
rate card, which is generated from the same constants
that charge the request, so it cannot quote a rate that will not be honoured.When there is nothing to quote
answer comes back empty, sources is empty, and message says so.
That is deliberate. An answer with no support is the one thing this endpoint
will not return, and inventing prose to fill the gap is exactly what every
other answer API does. Widen the filters or use a deeper tier so more pages get
read.