> ## Documentation Index
> Fetch the complete documentation index at: https://darwin.so/docs/llms.txt
> Use this file to discover all available pages before exploring further.

# Evaluate Search

> Measure ranking quality with representative intents and executable outcomes.

Evaluate Search against the intents your product actually receives. A good test set includes straightforward requests, ambiguous requests, hard constraints, unavailable supply, and cases that should return no acceptable option.

## Build an evaluation set

For every test case, record:

* the user intent and relevant context
* required and prohibited capability attributes
* acceptable capability IDs or a clear human relevance rubric
* whether an executable result should exist
* the capability revision used when the judgment was made

Keep evaluation labels separate from production ranking. Review them when capabilities or requirements change.

## Useful measures

| Measure               | What it tells you                                                  |
| --------------------- | ------------------------------------------------------------------ |
| Success at `k`        | Whether an acceptable capability appears in the first `k` results. |
| First executable rank | How quickly a user reaches an option they can act on.              |
| Constraint accuracy   | Whether filters and stated requirements are respected.             |
| Zero-result precision | Whether Search avoids presenting an invalid option.                |
| Ranking stability     | Whether unchanged requests avoid unexplained large shifts.         |

## Review failures

Classify failures as query ambiguity, missing capability metadata, incorrect filtering, stale availability, or ranking quality. Do not replace these labels with a fabricated confidence score.

Run the same set before changing query construction, filters, or result presentation. For end-to-end quality, continue the selected result through an [Act evaluation](/docs/features/act/evaluate).