vibe.review.llm¶

LLM client for document classification.

Provides both mock and real LLM clients for requirement classification. Mock client is used for testing and development; real clients integrate with Berget API or other OpenAI-compatible endpoints.

Design Note: Separate Implementation from vibe.llm_providers¶

This module intentionally provides its own LLM clients rather than reusing the providers from vibe.llm_providers (used by Assistant). The rationale:

1. Module Separation: Per the implementation plan (Section 1.7), Review is designed as a separate package (vibe-review) that depends on Core but avoids deep coupling. Core's llm_providers are tightly integrated with Assistant's conversation model (streaming, tool calling, multi-turn).

2. Different Use Cases:

3. Assistant: Interactive conversation with streaming responses, tool calling, multi-turn context, complex message history management

4. Review: Single-shot classification with structured JSON output, no streaming, no tools, no conversation history

5. Structured Output: Review requires strict JSON schema enforcement via the response_format parameter for deterministic classification results. This is a specific feature that Review uses extensively but Assistant doesn't need in the same way.

6. Simpler Interface: Review's classify() method is purpose-built for classification tasks: system prompt + user prompt → structured response. No need for the complexity of conversation management.

7. Dependency Footprint: This implementation uses httpx directly rather than the openai SDK, keeping Review's dependencies minimal.

If Review's needs grow closer to Assistant's (e.g., multi-turn classification discussions), consider refactoring to share infrastructure. For now, the ~100 lines of client code is simpler than adapting Core's providers.

See Also: - vibe.llm_providers.openai.OpenAIProvider (Core's provider for Assistant) - vibe.embedding_providers (shared embedding abstraction) - vibe.rerank_providers (shared reranking abstraction)

classify(system_prompt: str, user_prompt: str, response_schema: dict[str, Any]) -> ClassificationResponse

classify_structured(system_prompt: str, user_prompt: str, response_schema: dict[str, Any]) -> StructuredClassificationResponse

close() -> None

set_question_response(question_id: str, answer: str | bool | list[str] | None, confidence: str = 'M', reasoning: str = 'Mock reasoning', supporting_part_ids: list[str] | None = None, needs_user_input: bool = False) -> None

classify(system_prompt: str, user_prompt: str, response_schema: dict[str, Any]) -> ClassificationResponse

classify_structured(system_prompt: str, user_prompt: str, response_schema: dict[str, Any]) -> StructuredClassificationResponse

set_response(system_prompt: str, user_prompt: str, response: ClassificationResponse) -> None

set_response_for_text(text_contains: str, response: ClassificationResponse) -> None

clear_calls() -> None

assert_called() -> None

assert_call_count(expected: int) -> None

close() -> None

classify(system_prompt: str, user_prompt: str, response_schema: dict[str, Any]) -> ClassificationResponse

classify_structured(system_prompt: str, user_prompt: str, response_schema: dict[str, Any]) -> StructuredClassificationResponse

close() -> None

enable_prompt_logging(log_dir: str | Path | None = None) -> Path

disable_prompt_logging() -> None

configure_mock_llm(deterministic: bool = True, default_classification: str = 'Y', default_confidence: str = 'M', default_reasoning: str = 'Mock classification reasoning.', responses: dict[str, dict[str, Any]] | None = None) -> None

clear_mock_llm_config() -> None

set_mock_response(prompt_hash: str, response: ClassificationResponse) -> None

create_llm_client(provider: str = 'mock', **kwargs: object) -> BaseLLMClient