Fact search (
client.search.fact), deep search (client.search.deep), and
filtered search (client.search.filtered) are deprecated. Use automatic
search with mode fact instead of fact search, QA or automatic search with
mode agent instead of deep search, and automatic search instead of a new
filtered search. Their endpoints keep working until they are retired; the
ndi fact-search and ndi deep-search commands are removed from CLI 0.27.0.
See Search.
Automatic search, QA, and journal-entry testing return jobs.
Once a knowledge graph is built for the workspace, automatic search (both routes) and QA use it without any extra argument.
Automatic search is single-shot. It answers with a one-pass fact search, or with an agent that reads documents, queries tables, and can select documents from the metadata catalog. routing.final names the route (fact or agent; results stored earlier can read deep or filtered), output is that route’s result, and output_job_id owns receipts and cursors. An agent output carries its answer (a deep-search result) and, when it answered from the catalog, a catalog table of up to 200 rows whose next_cursor pages the rest through filtered search’s page form, which is not deprecated. Grounding ids for a fact route are on output. mode fact or agent runs that route with no routing step; routing.router.kind then reads requested.
Journal-entry testing is the one entry here that is not retrieval. Automatic search and QA retrieve and cite; je-testing performs a test — it writes its own SQL over the ingested ledger tables and publishes every statement as a receipt, so each figure in the answer traces to the query behind it. Reach for it when the deliverable is a tested criterion, not a cited passage.
QA is the conversation. Chain follow-ups with the previous turn’s answer.session_id. Sharing a session across a suite of unrelated questions is unsupported.
Automatic search, QA, and journal-entry testing plan Ground jobs into
grounding_job_ids on the answer: output for an automatic fact answer,
output.answer for the agent, answer for QA, and the result itself for
journal-entry testing. After the search result, poll each id with
client.jobs.wait (or GET /v1/jobs/{id}). Do not call
client.documents.ground for those evidences — that endpoint is for
standalone quotes, or for search jobs that predate the field.
See the search guide for runnable examples.