Yet Another LLM Client

An opinionated python wrapper for LLM calls. Supports multiple LLM providers:

• OpenAI
• Anthropic
• more to come...

Uses pydantic models to serialize LLM responses. Every response has to be serialized into a pydantic model.

Full async support.

Checking models

To verify which models in LLMModel are reachable with your current API keys:

cp .env.example .env  # fill in your API keys
uv run python scripts/check_models.py

Each model is called concurrently and results print as they complete.

Usage

Every call to the LLM returns some metadata. Metadata contains token usage, costs, model used and context messages. YALC supports 2 modes of operations for handling metadata.

Metadata return mode

Metadata is returned directly alongside the response as a tuple.

client = create_client(LLMModel.gpt_4o_mini)

result, metadata = await client.structured_response(
    JudgmentResult, messages
)

Advantages:

• Simple, no setup required
• Direct access to metadata at the call site

Disadvantages:

• Must handle metadata manually on every call
• Easy to forget or handle inconsistently across call sites

Strategy metadata mode

A metadata handler strategy is provided during client creation. The strategy is automatically invoked on every call when a context is passed. The provided context is used for any additional data that needs to be used when handling LLM call metadata.

# 1. Define your strategy
class LogStrategy(ClientMetadataStrategy[LLMLogContext]):
    def handle(self, call: ClientCall, context: LLMLogContext):
        print(f"Tokens: {call.input_tokens + call.output_tokens}")
        print(f"Cost: {call.input_tokens_cost + call.output_tokens_cost}")
        db.save(call.model_dump(), context.request_id)

# 2. Create client with the strategy
client = create_client(LLMModel.gpt_4o_mini, metadata_strategies=[LogStrategy()])

# 3. Pass context to trigger the strategy
result = await client.structured_response(
    JudgmentResult, messages, context=llm_log_context
)

Advantages:

• Metadata handling is set up once and applied consistently
• Call sites stay clean — no need to unpack or handle metadata each time

Disadvantages:

• More initial setup
• Metadata handling is implicit, which can be harder to trace