recipe/

Reproducible build recipe for the int8 ONNX model that ships inside the guild release binary (-tags=withembed).

The recipe is the source of truth. The model artifact is downstream: recipe/quantize.py produces it, .github/workflows/build-model.yml publishes it as a GitHub Release, and .model-version at the repo root pins the release tag that the binary release workflow consumes.

Files

• quantize.py: the two-step recipe (optimum-cli export, then onnxruntime quantize_dynamic with QInt8). Reads the upstream model name as a constant; writes outputs under workspace/models/.
• requirements.txt: pinned versions of optimum, onnxruntime, and the pieces the export pipeline pulls in.

Run locally

Validates the recipe end-to-end without going through CI. Useful when bumping requirements.txt or eyeballing the int8 SHA256 before publishing a model release.

python3 -m venv .venv
source .venv/bin/activate
pip install -r recipe/requirements.txt
python recipe/quantize.py

Outputs:

• workspace/models/bge-small-fp32/model.onnx (intermediate, ~127 MB)
• workspace/models/bge-small-fp32/vocab.txt
• workspace/models/bge-small-fp32/tokenizer.json
• workspace/models/bge-small-int8/model.onnx (final, ~33 MB)

Total run time: roughly one minute on a recent laptop, dominated by the HuggingFace download of the FP32 base model.

Relationship to .model-version and the build-model workflow

recipe/quantize.py                 (you are here)
        |
        v
.github/workflows/build-model.yml  (runs the recipe in CI, uploads
        |                           model.onnx, vocab.txt,
        |                           tokenizer.json, MANIFEST.txt as
        |                           model-v<semver> release assets)
        v
.model-version                     (semver string the binary release
        |                           workflow reads)
        v
.github/workflows/release.yml      (downloads model-v$VERSION assets,
                                    runs `make assets-model`, then
                                    goreleaser produces the embed
                                    binary)

Triggers for build-model.yml:

1. workflow_dispatch (manual) with optional version input. The maintainer runs this once after merging this quest to bootstrap the first model release.
2. push on recipe/** paths. Any change to this directory rebuilds the model.
3. schedule quarterly (cron 0 0 1 */3 *). Catches upstream BAAI changes and verifies the recipe still produces a healthy artifact.

The output of every model build is auditable via the MANIFEST.txt asset on each model-v* release: SHA256s of all files, optimum and onnxruntime versions used, BAAI source revision, build timestamp, and the workflow run URL.