Add default Rhesis embedding model and standardize terminology

[EmanueleDeRossi1]

Summary

This PR standardizes the terminology across the codebase from inconsistent "LLM/llm" naming to clear "language model" and "embedding model" terminology, and adds support for a default Rhesis-hosted embedding model.

Key Changes

1. Terminology Standardization

• Renamed get_model() → get_language_model() in SDK for clarity
• Changed model_type enum value from llm → language
• Updated all variable names from DEFAULT_GENERATION_MODEL → DEFAULT_LANGUAGE_MODEL_PROVIDER
• Consistent use of "language model" and "embedding model" throughout codebase
• Updated 73 files across backend, frontend, SDK, and tests

2. Default Rhesis Embedding Model

• Added RhesisEmbedder class in SDK with full implementation
• New endpoint: /generate/embedding (same behavior as /generate/content)
• Registered "rhesis" provider in embedding model factory
• Changed default embedding provider from OpenAI to Rhesis
• Backend now creates Rhesis Default Embedding model during organization initialization
• Stores both language_model_id and embedding_model_id in user settings
• Uses RHESIS_API_KEY and RHESIS_BASE_URL environment variables

3. Database Migration

• Added migration to update existing model_type values from "llm" to "language"

❗ Breaking Changes

Environment variable names change:

• DEFAULT_GENERATION_MODEL → DEFAULT_LANGUAGE_MODEL_PROVIDER
• DEFAULT_MODEL_NAME -> DEFAULT_LANGUAGE_MODEL_NAME

New variables to add:

• DEFAULT_EMBEDDING_MODEL_PROVIDER
• DEFAULT_EMBEDDING_MODEL_NAME

[peqy]

Critical issues to address before merge

1. Backend model_type values are inconsistent ("language" vs "language_model") across DB enum/migration, API schemas, and connection testing. This will cause 422s and/or make model connection tests fail.

2. SDK factory removed DEFAULT_PROVIDER but tests (and likely external users) still import it.

3. SDK embedding factory default provider changed to rhesis, but tests still assert OpenAI is used on the default path.

Found 6 issues (6 critical).

[peqy]

@EmanueleDeRossi1 yep, confirmed — the backend API schema/routers now consistently use model_type="language" (instead of language_model), which should unblock the earlier 422/“invalid model type” issues.

Remaining blockers from my earlier notes still look outstanding in this draft:

• sdk.models.factory still doesn’t define DEFAULT_PROVIDER (tests/imports still reference it)
• get_embedding_model() now defaults to rhesis, but the default-path unit tests still patch/assert OpenAIEmbedder (and will fail unless updated/mocked)

[peqy]

@EmanueleDeRossi1 confirmed — the backend is now consistently using model_type="language" (schema + router fallback + connection service), so that earlier mismatch should be resolved.

The other two blockers I flagged still remain (missing DEFAULT_PROVIDER backward-compat alias in sdk.models.factory, and the embedding-factory tests still expecting OpenAI on the default path).

[peqy]

@EmanueleDeRossi1 confirmed — with 561d187 the model_type handling in ModelConnectionService is now consistently "language", so the earlier “invalid model type” path should be resolved.

The two remaining draft blockers from my earlier notes still apply: sdk.models.factory missing the DEFAULT_PROVIDER backward-compat alias, and the embedding factory tests still expecting OpenAI on the default path even though get_embedding_model() now defaults to rhesis.

[peqy]

@EmanueleDeRossi1 confirmed — after 46fed56 I don’t see any remaining DEFAULT_PROVIDER imports/usages in the repo, so this is resolved on our side.

One related follow-up: get_embedding_model() still documents the default as OpenAI and tests/sdk/models/test_embedder_factory.py::test_get_embedding_model_minimal_defaults still patches OpenAIEmbedder, but the default provider is now rhesis (DEFAULT_EMBEDDING_MODEL_PROVIDER). Likely needs aligning to avoid CI failures.