Summary

This PR removes hardcoded embedding assumptions (provider wiring, model name, and 1536 dimensions) so self-hosted deployments can use OpenAI-compatible embedding providers (e.g. Ollama) without source overrides.

Root Cause

Current code hardcodes OpenAI-style defaults in multiple places:

• src/embedding_client.py
  • openrouter model hardcoded to openai/text-embedding-3-small
  • openai model hardcoded to text-embedding-3-small
  • Gemini requests hardcode output_dimensionality=1536
  • no explicit custom path for embedding provider even though non-embedding LLM config uses custom
• src/models.py
  • Document.embedding and MessageEmbedding.embedding use Vector(1536) directly

This forces downstream operators to mount patched embedding_client.py, models.py, and config.py into the container when using embedding models with different dimensions (e.g. qwen3-embedding:4b -> 2560 dims).

Changes

1. Add configurable embedding model + provider compatibility in settings

• src/config.py
  • LLMSettings.EMBEDDING_PROVIDER now accepts custom (keeps openrouter alias)
  • add LLMSettings.EMBEDDING_MODEL: str | None = None

2. Align embedding provider handling with LLM configuration pattern

• src/embedding_client.py
  • Support custom embedding provider using:
    • LLM_OPENAI_COMPATIBLE_API_KEY
    • LLM_OPENAI_COMPATIBLE_BASE_URL
  • Keep openrouter as compatible alias (default base URL fallback)
  • Validate unsupported embedding provider values with explicit error
  • Use EMBEDDING_MODEL override for gemini/openrouter/custom/openai

3. Remove hardcoded Gemini dimensionality

• Stop forcing output_dimensionality=1536 in all Gemini embedding calls
• Let provider/model defaults apply

4. Make ORM vector dimensions follow config

• src/models.py
  • Vector(settings.VECTOR_STORE.DIMENSIONS) for:
    • MessageEmbedding.embedding
    • Document.embedding

5. Add tests (including regression coverage)

• tests/test_embedding_configuration.py
  • custom provider path (key/base URL/model wiring)
  • custom requires base URL (error case)
  • openrouter alias behavior
  • openai behavior
  • invalid provider rejection
  • regression: gemini embed calls no longer pass config.output_dimensionality
  • wrapper singleton resolves custom provider API key path
  • regression: ORM embedding dims follow settings.VECTOR_STORE.DIMENSIONS

Why this is safe

• Backward-compatible defaults remain for existing users.
• Existing 1536 deployments continue to work unchanged.
• Self-hosted operators can now align:
  • LLM_EMBEDDING_PROVIDER=custom
  • LLM_EMBEDDING_MODEL=<provider model>
  • VECTOR_STORE_DIMENSIONS=<matching dim>
    without code mounts/overrides.

Validation

• Ruff:
  • uv run ruff check src/config.py src/embedding_client.py src/models.py tests/test_embedding_configuration.py
  • uv run ruff format --check src/config.py src/embedding_client.py src/models.py tests/test_embedding_configuration.py
• Targeted tests (isolated from repo conftest):
  • uv run pytest --noconftest tests/test_embedding_configuration.py -q
  • Result: 8 passed

Operator Impact

This should allow removing runtime code overrides like:

• mounted embedding_client.py
• mounted models.py
• mounted config.py
  from docker-compose in self-hosted setups.