Fix: the embedding model default breaks one deployment or the other

[adamjohnwright]

Current main fails on the first query for any Reactome deployment that does not set EMBEDDING_MODEL.

What is actually wrong

graph.py defaulted to the literal "bge-m3". Against api.openai.com:

The model `bge-m3` does not exist or you do not have access to it. (404)

But bge-m3 is not a bad model. It is the correct one for Plant Reactome, which serves it from a self-hosted OpenAI-compatible endpoint via OPENAI_BASE_URL — the bundle is openai/bge-m3/plantreactome/Release68. It is wrong only as a default, because every bundle published for Reactome uses text-embedding-3-large.

So hardcoding either model breaks the other deployment. My first attempt at this fix hardcoded the Reactome bundle and would have broken Plant Reactome; this corrects that.

The fix

The bundle path records the model that built it, which makes it the source of truth rather than a constant kept in sync by hand. resolve_embedding_model():

1. reads the model from all installed bundles
2. reports it as a misconfiguration when they disagree — AgentGraph builds one embedding shared by every profile, so bundles built with different models cannot all be served
3. honours EMBEDDING_MODEL as an override, and logs an error when the override disagrees with the bundle, since that gives meaningless retrieval rather than an obvious failure
4. falls back to text-embedding-3-large only when nothing is installed

Verified against the real Release95 bundle: resolves to text-embedding-3-large.

Where it came from, and a process point

The default arrived in a8aa04c — "Initial commit of adding PlantReactome profile", Justin Elser, 2026-08-19.

That branch has no pull request, in any state. It reached main today through the integration PR #167 without ever having had a review of its own, and it carried a default that breaks the primary deployment. Neither the lint nor the type gate can see this: it is a runtime configuration mismatch, and nothing exercised it.

Plant Reactome is unaffected in every configuration, verified:

bundle installed	EMBEDDING_MODEL	resolves to
openai/bge-m3/plantreactome/Release68	unset	bge-m3
same	bge-m3	bge-m3
same	something else	the override, with a loud error

It is in fact more robust than before: it previously relied on a hardcoded default that happened to match, so changing that default for Reactome would have silently broken Plant Reactome. It now derives from the bundle actually installed.

Also

EmbeddingEnvironment.get_model raises KeyError for an uninstalled database while its sibling get_dir returns None. Not changed here — it is a shared API — but worth knowing.

Tests

Seven: Reactome resolves to text-embedding-3-large, Plant Reactome to bge-m3, override honoured, override-disagrees reported, mixed bundles reported, nothing-installed falls back, and one asserting the default is specifically never bge-m3 when a Reactome bundle is installed.

Deployed instances

beta.reactome.org/chat and production both run image e398a37, which predates this code, so neither is affected.

🤖 Generated with Claude Code