Skip to main content

Evaluations

An Evaluation ties together the core building blocks of a search relevance assessment:

Once you create an evaluation, you can run it repeatedly to track how your search relevance changes over time.

Creating an Evaluation​

In the UI​

  1. Navigate to Evaluations and click Create Evaluation
  2. Enter a Name and optional Description
  3. Select the Endpoint, Query Set, and Query Template to use
  4. Click Create

Using the API​

curl -X POST "https://${RELEVAL_HOST}/api/v1/evaluations" \
-H 'Content-Type: application/json' \
-H "Authorization: Bearer ${TOKEN}" \
--data @- <<EOF
{
  "name": "Product Search v2",
  "description": "Evaluate new ranking model",
  "endpoint_id": "${ENDPOINT_ID}",
  "query_set_id": "${QUERY_SET_ID}",
  "query_template_id": "${QUERY_TEMPLATE_ID}"
}
EOF

Evaluation Workflow​

The typical workflow for evaluating search relevance:

  1. Create an Evaluation: define what you're evaluating
  2. Create an Evaluation Run: choose the relevance scale and metrics
  3. Start the run: Releval executes each query against your endpoint and collects results
  4. Judge the results: rate how relevant each returned candidate is
  5. Review metrics: Releval automatically calculates metrics from your judgments
  6. Iterate: create new runs to track improvements as you adjust your search configuration

Comparing runs​

Because each evaluation can have multiple runs, you can track how metrics evolve:

  • Run before and after changing a ranking model
  • Run against different endpoints (production vs. staging)
  • Run with different query templates to compare search strategies

Each run preserves its results and metrics independently, giving you a historical record of search quality. To dig into a specific change, compare two runs head-to-head: the comparison reports per-query relevance metric deltas, how much the result lists changed (RBO, Jaccard, and candidate churn, independent of judgments), and execution-time differences, down to a side-by-side diff of the candidates each run returned.