Skip to content
GOCUX
Documentation
GOCUX / Documentationsearch

Evaluate search with a gold set

Measure whether project retrieval finds the evidence your workflow needs.

Measure whether project retrieval finds the evidence your workflow needs.

Practical procedure

  1. Create a small set of real questions and manually identify acceptable supporting passages. Include ambiguous terms, outdated documents and questions with no answer in the selected corpus.
  2. Run each query against a fixed document set and record the top passages, source versions and result limit. Count useful supporting hits separately from superficially related hits.
  3. Review failures as missing source material, extraction loss, query mismatch or ranking problems. Change one variable and repeat the same questions before expanding the evaluation.

Search the project corpus

Knowledge holds document text and provenance inside a project. Create a source, upload a document into that source, and inspect the extracted text before relying on a search result. Full-text retrieval is available after ingestion; semantic retrieval contributes when the embedding configuration and backfill are available. This operation searches the uploaded corpus, rather than fetching arbitrary web pages.

Contract example

Developer examples below describe routes on an authorized, configured GoCux control-plane base URL. They are request shapes or operation references, not claims that a particular public hostname exposes the API. Use the workspace for the corresponding human workflow; availability and permissions depend on the actual deployment.

POST /knowledge/search — example request body

{
  "projectId": "project-id",
  "query": "renewal conditions effective date",
  "limit": 10
}

Inspect the result

The response contains the query and a hits array. Use the returned chunk provenance to trace each hit to its document and source. Compare the actual passage with the question before treating it as evidence. Search ranking helps find material; it does not itself establish that a claim is correct.

What to keep explicit

Report retrieval coverage, latency and corpus identity together. A fixture evaluation does not establish production capacity or the correctness of a generated answer.

Handle incomplete and refused work

projectId must identify an existing project; the query is 1–1,000 characters and limit is 1–50, default 10. An empty hits array is a valid result. A 503 is an unavailable store and must remain visibly different from an empty result. Semantic unavailability must not be described as a complete search outage when full-text search still works.

Bring this workflow into your workspace.

Set the scope. Review the evidence. Approve the steps that need it.

Open the workspace

Keep exploring