Skip to content
CogniYukti

Finding it

A search that returns everything has returned nothing.

Rank the whole corpus and the bottom of the list is noise wearing a score. The real hits are still at the top — and nobody scrolls far enough to be sure, so they stop trusting it and ask a colleague instead.

The fix is a bar for being included at all, separate from how things are ordered.

Two decisions, not one

Whether a document belongs in the answer, and where it sits in it.

A document has to clear a bar: it contains the words, or it is close enough on spelling, or it is close enough in meaning. Otherwise it is left out rather than ranked last.

The reason is written in the code — ranking alone returned every row that passed the workspace filter at the bottom of the list, drowning the real hits in things that merely existed.

Askedwhy does the import keep failing after we changed the schema?
  • Rebuilding the index after a schema changethe phrase, almost exactly
  • Import failures and what causes themdifferent words, same subject
  • Setting up your first connectoradjacent, and not an answer
  • Expense approval thresholdsnothing to do with it
The third row is the interesting one. It is related and it is not an answer, so it is left out rather than ranked last — and when nothing clears the bar, the honest response is that there is no evidence.
Illustrative
  • The meaning threshold was retuned, and the history is kept

    The original value was chosen against direct-phrase queries and dropped every paraphrase. The observed distributions are recorded alongside the new one — the sort of note that stops somebody "tidying" it back in a year.

  • Spelling tolerance exists for a named reason

    Partial words and typos the word index misses entirely.

  • Degradation is reported rather than silent

    When the meaning half is unavailable the response says so, because hiding it once cost somebody an hour of thinking the search was broken.

  • An empty query still works

    Falling back to what is verified and recent rather than erroring.

What you are allowed to see

Six sources, filtered by what your workspace actually has.

Platform-wide material is indexed for everyone, so the filtering has to happen at the point of search. A document carries the module it belongs to, and the filter admits one that carries no such marking at all — so getting-started material stays visible to a workspace with nothing switched on, while a sales-only workspace never turns up an HR policy.

Internal and general visibility is a single permission, and every search is recorded with how many results it returned — which feeds the report of the most common searches that came back empty.

Does it search records as well as knowledge?

No — and the product keeps them deliberately separate, with two different shortcuts. Records are found by one search, knowledge by this one. Two different questions.

Can access be restricted per category?

No. Categories organise; they do not control access. Visibility is one permission across the workspace.

Are drafts findable?

A draft's body is empty until it is published, so a search of the text will not find it — but treat titles as potentially visible rather than assuming otherwise.

Finding it

Keeping it good