chops-search

Reindex on every push

The index should be built by CI, not committed: gitignore static/search/, make the pipeline the single producer, and every content push reindexes automatically. On a content-only push the steady-state cost is a few seconds.

The shape

Three cached inputs, one build, one gate, then your existing deploy:

  1. The chops-search binary, cached against the pinned version.
  2. The model, cached against the committed model.lock.json.
  3. chops-search build into static/search/, before your SSG build so the output gets copied into the deployable site.
  4. chops-search eval --fail-under as the quality gate.

GitHub Actions

.github/workflows/deploy.yml (excerpt) 
env:
  CHOPS_SEARCH_VERSION: "0.2.10"

steps:
  # ---- chops-search binary, cached on the pinned version ----
  - name: Restore chops-search
    id: chops-bin
    uses: actions/cache/restore@v4
    with:
      path: ~/.cargo/bin/chops-search
      key: chops-search-${{ runner.os }}-${{ env.CHOPS_SEARCH_VERSION }}

  - name: Install chops-search
    if: steps.chops-bin.outputs.cache-hit != 'true'
    run: cargo install chops-search --version "$CHOPS_SEARCH_VERSION" --locked

  - name: Save chops-search
    if: always() && steps.chops-bin.outputs.cache-hit != 'true'
    uses: actions/cache/save@v4
    with:
      path: ~/.cargo/bin/chops-search
      key: chops-search-${{ runner.os }}-${{ env.CHOPS_SEARCH_VERSION }}

  # ---- model, cached on the committed lockfile ----
  - name: Restore model
    id: model
    uses: actions/cache/restore@v4
    with:
      path: .chops-search/model
      key: model-${{ hashFiles('.chops-search/model.lock.json') }}

  - name: Fetch model at locked revision
    if: steps.model.outputs.cache-hit != 'true'
    run: |
      REV=$(jq -r .revision .chops-search/model.lock.json)
      chops-search model fetch --revision "$REV"

  - name: Verify model against lockfile
    run: chops-search model verify

  - name: Save model
    if: always() && steps.model.outputs.cache-hit != 'true'
    uses: actions/cache/save@v4
    with:
      path: .chops-search/model
      key: model-${{ hashFiles('.chops-search/model.lock.json') }}

  # ---- index + runtime, then gate ----
  - name: Build search index
    run: chops-search build

  - name: Eval search quality
    if: hashFiles('fixtures/queries.toml') != ''
    run: chops-search eval --fail-under 0.85

  # ---- then your existing SSG build and deploy, unchanged ----

Why each piece is shaped that way

The restore/save split with if: always(). The combined actions/cache saves in a post step that never runs if the job fails first, so a red build after the 30 MB model fetch would throw the download away. Splitting restore and save keeps failed builds from costing the next run.

--revision from the lockfile, then verify. model fetch on its own re-resolves upstream and rewrites the lock, which CI must never do. Passing the locked revision keeps the build reproducible, and running verify unconditionally (even on cache hit) means a corrupted or evicted-and-restored cache can't feed the build silently wrong bytes.

The full build, not --no-runtime. Since static/search/ is gitignored, every CI run must produce the runtime too. The runtime is embedded in the pinned binary, so runtime and artifact format can't drift.

The gate is conditional on the fixture existing. The workflow is paste-able before you've written a query set, and the gate switches itself on the moment fixtures/queries.toml merges. Writing one is worth an hour.