Skip to main content
One search box, two behaviours. Type for live name and slug suggestions; press Enter to search by meaning. It is on both the Projects and Experiments pages, and it is available even when you have no experiments — project-direct runs are searchable too.

​
Filter grammar

Operators are parsed deterministically. Whatever prose remains after parsing goes to semantic retrieval. Supported operators: entity type, lifecycle status, relative dates (today, yesterday, last week, last N days), project:, owner:, tag:, after: and before:, config.<key> comparisons, and metric.<key> comparisons.

​
Chips

Every interpretation is shown as a removable chip, so you can correct the parse without rewriting the query. Explicit UI choices — the entity-type control, the workspace — take precedence over anything inferred.
Unrecognised or ambiguous wording stays semantic text, or produces an actionable validation message. It never silently invents a metric key, threshold, aggregation or owner.

​
Metric comparisons

A metric predicate uses the latest scalar for an exact metric key with unambiguous series dimensions. An absent or ambiguous measurement does not match — it is not treated as zero and not treated as passing. Structured predicates apply before the candidate limit, so a pure filter query enumerates matching rows deterministically. For conceptual queries the response marks its window and any truncation honestly — a filter applied after a bounded engine window does not guarantee exhaustive recall, and the response does not pretend otherwise.

​
Results

Results navigate directly to the project, experiment or run, showing type, the current breadcrumb, status and a matched excerpt. Matching runs expand under their experiment; repeated siblings collapse after three, without ever hiding a confirmed exact identifier. The All / Projects / Experiments / Runs controls are server filters, not filters over what is already on screen. A confirmed complete name, slug or UUID match is pinned to the top. Other evidence is deduplicated and combined by reciprocal-rank fusion over channel ranks — never by adding raw scores.
An empty result list never claims exhaustive absence when retrieval is degraded or capped. If the semantic channel is unavailable, the response says so rather than reporting zero hits as a finding.

​
Workspace scope

Search respects the selected workspace. Choose All workspaces for cross-workspace discovery — the most common “search is broken” report is a project filed on a different shelf. Changing workspace or query aborts obsolete requests, so an answer for inputs you have left is never displayed.

​
The knowledge index

Beyond entities, one index covers documents: captured sessions, notes, papers, connected GitHub repositories, imported W&B and Benchling content, and artifact text. Two channels answer every query — an exact channel that matches identifiers, names and paths, and a semantic channel that matches meaning — sectioned by source with per-result provenance.
A literal identifier is the strongest query you can write. An issue number, a run slug, a file path or a hash goes straight through the exact channel. Phrasing matters much less than people expect; identifiers matter much more.
Search previews position their window on the chunk that matched the query, not on the first N characters of the document — which on many documents is a metadata header rather than anything about your question.

​
For agents

See the MCP reference.

​
When the engine is off

A self-hosted install can run without the knowledge engine. Search then returns state: "partial" with SQL-exact results only — matching projects, experiments and artifacts, but not run names, and with no semantic channel. That state is reported, not hidden. See self-hosting.