pynydus.llm.client

Instructor-based LLM client factory and completions. Spec §9.3.

Provides a vendor-agnostic interface for LLM calls with structured Pydantic output. Each LLMTierConfig specifies its provider, model, and API key: no defaults, no fallbacks.

Required dependency:

• instructor >= 1.0

Module Contents

Functions

create_client	Create an Instructor client for the given tier config.
create_completion	Run a structured LLM completion and return a validated Pydantic model.

Data

T
_logger

API

pynydus.llm.client.T

‘TypeVar(…)’

pynydus.llm.client.create_client(tier: pynydus.llm.models.LLMTierConfig) → instructor.Instructor

Create an Instructor client for the given tier config.

Uses instructor.from_provider which supports Anthropic, OpenAI, Mistral, Cohere, Google, and other vendors via a "provider/model" string.

Args: tier: Provider, model, and API key (all required).

Returns: Configured Instructor client.

pynydus.llm.client._logger

‘getLogger(…)’

pynydus.llm.client.create_completion(tier: pynydus.llm.models.LLMTierConfig, messages: list[dict[str, str]], response_model: type[pynydus.llm.client.T], *, max_retries: int = 3, log: list[dict] | None = None) → pynydus.llm.client.T | None

Run a structured LLM completion and return a validated Pydantic model.

Returns None (instead of raising) when the LLM is unavailable so the pipeline can fall back to deterministic-only mode.

Args: tier: LLM tier config to use for this call. messages: Chat messages as [{"role": ..., "content": ...}, ...]. response_model: Pydantic BaseModel subclass. Instructor validates output against this schema (with retries on validation failure). max_retries: Max retries when validation fails. log: If set, appends an {"type": "llm_call", ...} entry with timing, provider, model, and optional token usage.

Returns: A validated instance of response_model, or None on failure.