Function calling¶

Fairmark generates tool definitions for any LLM that supports function calling. They are built from the engine's JSON Schemas, so they always match the implementation.

Provider	Definitions	Format
Anthropic Messages API	`fm.tools.anthropic()`	`name`, `description`, `input_schema`
OpenAI and OpenAI-compatible Chat Completions	`fm.tools.openai(portable=False)`	`{"type": "function", "function": {...}}`
OpenAI Responses API	`fm.tools.openai_responses()`	`{"type": "function", "name": ..., "parameters": ...}`
Google Gemini	`fm.tools.gemini()`	`name`, `description`, `parametersJsonSchema`
By provider name	`fm.tools.for_provider("ollama")`	the right format for that provider
Anything else	`fm.tools.definitions(portable=False)`	`name`, `description`, `parameters` (JSON Schema)

Supported providers and complete examples: LLM providers.

Executing calls¶

result = fm.tools.dispatch(name, arguments)   # arguments: dict or JSON string

dispatch never raises on bad input. It returns {"error": {"code", "message"}}, which you send back to the model as the tool result so it can fix its parameters.

Tool names¶

Method dcf.fcff becomes tool fairmark_dcf_fcff. Use fm.tools.method_id(name) to map back.

Schemas the model sees¶

Schemas are inlined (no $ref) for compatibility across providers. Decimal fields accept numbers or strings; descriptions state units explicitly ("decimal fraction, 0.09 = 9%"), which is what makes models extract rates correctly. The portable variant (portable=True, used for Gemini and local open-weight runtimes) declares decimals as strings, turns optional values (X or null) into plain X (an optional field is simply not required), and removes pattern and additionalProperties. Dates are ISO 8601 strings ("2026-06-30") in both variants.