Improve the query before repeating the call: Kimi Search Pro filters and relevance scores

When a Kimi Search Pro response misses the evidence you need, repeating the request with synonyms is not the most useful next step. First decide what was underspecified: the entity, the period or version, the source scope, or the question itself.

The reason is concrete: Kimi’s web-search best-practices guide says text_query is searched as submitted, without automatic rewriting, completion or expansion. The guide also states that calls are billed per request. Query construction is therefore a decision to make before another paid call—not something to assume the service will repair.

This is an explanation of the documented interface, not a release announcement. The source was captured on September 18, 2026; its publication date is not established.

Refine one information need, not just its wording

Kimi recommends making three things explicit:

  • Entity: name the product, organization, person or event rather than a broad category.
  • Time or version: specify the year, quarter or version relevant to the question.
  • Angle or source type: say whether you need official documentation, an announcement, reporting or discussion.

For an illustrative documentation search, “search filters” is less specific than “Kimi Search Pro sites time_window official documentation.” If the missing information is how site restrictions interact with date restrictions, that is a related follow-up worth expressing in the same query. A separate question about another product’s subscription plans is not.

The guide recommends consolidating follow-ups about the same information need and describes substantial overlap between near-identical queries. That is the documentation’s rationale for avoiding synonymous retries, not a measured overlap rate or a guarantee that one longer query will recover every useful source. A more complete question is the goal; a bag of unrelated questions is not.

Ranking hints and scope restrictions are different controls

What you need Control documented by Kimi What not to assume
Prefer a source type Include wording such as “official documentation” in text_query That wording is not the same as a site restriction.
Restrict the source sites Set sites, with at most five sites, combined with OR Adding more sites does not require each result to match all of them.
Express the relevant year or version Include it in text_query; time words affect search and ranking A year in the query is not a publication-date filter.
Restrict when pages were published Set time_window start/end values; supported date formats are YYYY, YYYY-MM, and YYYY-MM-DD Publication time is not necessarily the date of the event discussed.

My recommendation is to use restrictions only when they express a real requirement. If only approved official sites are acceptable, sites makes that requirement explicit. If the question concerns an older event but later retrospective analysis is useful, a narrow publication window could exclude relevant evidence. Query time words and time_window can be combined, but they solve different problems.

A higher score on the next call is not proof of improvement

The guide defines chunks[].score as relevance between a chunk and the query. Its warning is important: scores have relative meaning only within the same query; do not compare absolute values across queries.

That rules out using a larger score after a rewrite as evidence that the second search is better. It also gives no basis for a universal cutoff such as “accept every chunk above this number” across all requests. Nor should a relevance score be read as a probability that a passage is factually correct.

For a web-grounded answer, inspect whether the passage actually supports the requested claim and retain its source link. Relevance to a question and support for an answer are different judgments.

A practical check before the next request

For an agent or RAG integration, I would use this small diagnostic sequence:

  1. Write down the missing evidence. “Need the documented site-count limit” is actionable; “results look weak” is not.
  2. Change the request for a reason. Add an entity, version or source qualifier; apply or relax a scope restriction when the information need warrants it. Do not merely swap synonyms.
  3. Keep the comparison reproducible. Record the original and revised query, site restrictions, date window and returned source links.
  4. Judge the revision by the missing evidence, not the score delta. Did it return a passage that answers the unresolved question? If not, identify a different source angle rather than treating another similar call as progress.

This is a suggested evaluation procedure, not a tested recall improvement or a quantified cost saving. The useful distinction is simpler: another call should pursue a defined evidence gap. Query-relative scores can help interpret that call’s chunks, but they cannot certify that the next call was worth making.

I would separate repairing one search from evaluating a query-rewrite policy. Finding the missing passage is a useful stopping point for that request, but it does not establish that the rewrite strategy should become the agent’s default.

For the second decision, I would use a small held-out set of information needs with acceptance criteria written before seeing the results. For example: “Find an official passage stating the site-count limit,” rather than “return a relevant-looking documentation page.” Compare the original and revised queries while keeping sites and time_window unchanged; test filter changes separately. Otherwise, a successful result would not tell you whether the wording or the restricted search scope helped.

I would hide which query produced each passage during judging and record three outcomes: the required evidence was found, only partial evidence was found, or the gap remains unresolved. Also count the calls used, including unsuccessful retries. A policy that occasionally finds better evidence after many attempts presents a different trade-off from one that does so on the first call.

This is a proposed evaluation design, not a measured result for Kimi Search Pro. It builds on the guide’s warning that scores are query-relative: evaluate the rewrite policy against fixed evidence requirements, not higher scores—or acceptance criteria relaxed after seeing what the search returned.