SchemaGate 0.1.44 chooses its embedder automatically—and turns an optional extra into runtime policy

SchemaGate 0.1.44 changes Catalog() so the library selects its embedding path from the installed environment rather than requiring every user to make that choice manually. A base installation continues to use the project’s hashed n-gram vectorizer; an installation with the huggingface extra automatically uses all-MiniLM-L6-v2. The release keeps SCHEMAGATE_AUTO_EMBEDDER=0 as an escape hatch for operators who want the hashed path even when the sentence-model dependency is present. (release notes)

The measured gain is real but narrow

The project reports 90 correct retrievals out of 98 questions for the hashed embedder and 93 out of 98 for MiniLM across six bundled schemas, with no schema descriptions supplied. All three additional successes came from the commerce schema’s conversationally phrased questions, where the score moved from 15/18 to 18/18. The author explicitly characterizes this as a three-question overall gain rather than a transformation across the whole suite. (implementation commit)

That distinction matters for deployment decisions. The base path remains deliberately small, offline and deterministic, while the sentence-model path is activated only for users who have already installed the heavier Hugging Face stack. In other words, 0.1.44 does not impose a model download on every installation; it uses the dependency set as a signal that the operator has accepted that trade-off. (release notes)

Installed packages now affect retrieval behavior

The convenience also creates a configuration boundary. Two hosts running the same application code can choose different embedders if only one has the Hugging Face extra installed. The project pins its own test suite to the hashed implementation so developers with different extras do not silently run different tests, and the commit notes that its JavaScript twin has no sentence-model path. (implementation commit)

Persisted indexes deserve particular care. The author says an upgraded persisted index can encounter a vector-dimension mismatch when automatic selection changes the embedder, and 0.1.44 updates that error to point operators toward the opt-out variable. Teams should therefore pin the intended path explicitly during rollout and test existing indexes before allowing dependency changes to select a new model. (implementation commit)

The release also reports a live Oracle check on a 30-million-row schema. It describes ten complex questions, then reports 9/9 answered and run, with eight reaching the full required join depth; three further questions whose requested relationships were absent were declined rather than answered with invented columns. The release does not explain the ten-versus-nine denominator, and these are project-reported results rather than an independent benchmark, but the check still exercises a useful failure mode alongside retrieval accuracy. (release notes)

For practitioners, the important change is less “sentence embeddings win” than “environment composition is now policy.” The upgrade removes a setup decision for new users, but production deployments should make that decision visible in dependency locks, environment configuration and index-migration tests.