LLM¶

Provider connections and named, concretely-configured models for LLM/embedding backends. Peer of the Resources domain: ProviderConfig plays the same role as ConnectionConfig, and ModelConfig plays the same role as DatabaseConfig.

ProviderConfig¶

A concrete LLM/embedding provider connection: provider key, base URL, API key. Stored in [providers.<name>].

Bases: BaseModel

Concrete connection to one LLM provider: which one, and how to reach it.

Peer of :class:~oa_configurator.domains.resources.schema.ConnectionConfig for LLM/embedding backends instead of databases. Referenced by :attr:ModelConfig.provider. Each entry under [providers] in config.toml maps to one instance of this model.

API keys are stored in plaintext for now, the same documented limitation ConnectionConfig.password already carries; secret management support is planned for a future release for both.

resolve ¶

resolve(name: str) -> ResolvedProvider

Resolve this provider to a concrete, backend-ready connection.

ModelConfig¶

A named, reusable, concretely-configured model, served through a provider. Stored in [models.<name>].

Bases: BaseModel

A named, reusable, concretely-configured model.

Peer of :class:~oa_configurator.domains.resources.schema.DatabaseConfig for LLM/embedding backends instead of databases. The unit that consuming packages reference by name (e.g. a package's embedding_model_name field just names an entry here). Each entry under [models] in config.toml maps to one instance of this model.

resolve ¶

resolve(name: str, stack: StackConfig) -> ResolvedModel

Resolve this model to a concrete, backend-ready configuration.

stack must already have passed :meth:StackConfig.validate_references, so self.provider is guaranteed to exist in stack.providers.

validate_embedding_configuration ¶

validate_embedding_configuration() -> ModelConfig

Reject embedding dimensions on models without embedding support.

Resolved types¶

ProviderConfig.resolve() and ModelConfig.resolve() produce these. Resolver.resolve_provider()/resolve_model() are thin wrappers around the same two methods.

Concrete LLM provider connection, ready to be served through.

Attributes:

Name	Type	Description
`name`	`str`	Logical name of the provider as declared in the config.
`provider`	`str`	Provider key, e.g. `'ollama'`, `'llamacpp'`, `'openai'`.
`base_url`	`(str, optional)`	Resolved base URL for this deployment.
`api_key`	`(str, optional)`	Resolved API key for this deployment.

Concrete, backend-agnostic model configuration. No explicit methods as it is just a data struct that can be used by the consuming package to construct a backend-specific model handle.

Attributes:

Name	Type	Description
`name`	`str`	Logical name of the model as declared in the config.
`provider`	`ResolvedProvider`	Resolved provider this model is served through.
`model`	`str`	Model name or identifier passed to the provider.
`embedding_dim`	`(int, optional)`	Embedding dimension override, or None to let the provider's own discovery determine it.
`document_prefix`	`(str, optional)`	Prefix prepended to document/passage text before embedding, for asymmetric embedding models.
`query_prefix`	`(str, optional)`	Prefix prepended to query text before embedding, for asymmetric embedding models.
`embeddings`	`bool`	Whether this specific model supports the embeddings endpoint.
`tool_use`	`bool`	Whether this specific model supports tool/function calling.
`structured_output`	`bool`	Whether this specific model supports structured (schema-constrained) output.
`extended_thinking`	`bool`	Whether this specific model supports reasoning/extended-thinking output.
`configuration`	`dict[str, Any]`	Free-form per-model knobs (max_tokens, temperature, and so on) with no dedicated field.