Optional real semantic model tests

The default test suite uses deterministic fixtures because CI must remain network-free, reproducible, and small. A separate promotion test downloads two revision-pinned Apache-2.0 MiniLM ONNX models:

  • cross-encoder/ms-marco-MiniLM-L6-v2 exercises P12 reranking.
  • sentence-transformers/all-MiniLM-L6-v2 exercises P13 embeddings.

The arm64 installer selects the roughly 23 MB quantized ONNX artifact for each model. Other architectures use the portable ONNX artifact. Model weights, tokenizer files, virtual environments, and reports live under the ignored target/real-semantic-models directory. They are not included in source or release packages.

Install and run the local CPU smoke test explicitly:

scripts/install-real-semantic-models.sh
scripts/run-real-semantic-model-smoke.sh

After installing pgContext into a local PostgreSQL 17 or 18 installation, run the database-bound promotion smoke separately:

PG_MAJOR=17 scripts/run-real-semantic-postgres-smoke.sh

It creates an isolated temporary PostgreSQL cluster, prepares a P12 envelope, scores it with the real cross-encoder, and finalizes the response through PostgreSQL. It also embeds source-linked chunks, builds a 384-dimension pgcontext_hnsw index, retrieves the relevant occurrence, and verifies that its citation object survived. The temporary cluster is removed on exit.

The installer requires uv and Python 3.12. It verifies exact model revisions and records each downloaded ONNX artifact’s byte length and SHA-256 digest in target/real-semantic-models/installed-models.json. The smoke report records the models, revisions, digests, platform, cold-start time, warm inference time, RSS, rerank scores, embedding cosine scores, and embedding dimensions.

The adapter can also score one rerank_envelope_v3 object from standard input:

target/real-semantic-models/venv/bin/python \
  tools/real-semantic-models/semantic_models.py rerank < envelope.json

Register or prepare that envelope with model name cross-encoder/ms-marco-MiniLM-L6-v2 and numeric model revision 1. The installed manifest binds that wire identity to the exact Hugging Face revision and artifact digest. The adapter rejects any mismatched identity instead of echoing it in a response.

For source-linked chunks, embed accepts objects shaped as {"chunks":[{"occurrence_id":1,"citation":{...},"text":"..."}]}. Its output preserves the occurrence and citation objects beside a normalized 384-dimension embedding. This is a test adapter, not an automatic P13 database job runner.

These smokes are initial real-model promotion checks, not Stable certification. Stable promotion still needs a retained held-out quality set, source/ACL/RLS churn, timeout and cancellation tests, hosted CPU/platform evidence, and a decision about whether a real inference runtime belongs inside pgcontext-worker or remains an application-managed adapter. P13 also still needs a production embedding-job integration; this smoke does not replace the automatic chunking schema’s deterministic fixture embedding.