Interface ILlmFunction

Coerce pre-parsed arguments to match expected schema types.

Use this only when the SDK (e.g., LangChain, Vercel AI, MCP) already parses JSON internally. For raw JSON strings from LLM output, use parse instead — it handles both lenient parsing and type coercion in one step.

LLMs often return values with incorrect types even after parsing:

• "42" → 42 (when schema expects number)
• "true" → true (when schema expects boolean)
• "null" → null (when schema expects null)
• "{...}" → {...} (when schema expects object)
• "[...]" → [...] (when schema expects array)

This function recursively coerces these double-stringified values based on the parameters schema.

Type validation is NOT performed — use validate after coercion.

Whether this function is deprecated.

When true, indicates the function should no longer be used. LLMs may still invoke deprecated functions but should prefer non-deprecated alternatives when available.

Human-readable function description.

Explains what the function does, when to use it, and any important considerations. This description is crucial for LLMs to understand when to invoke this function. Extracted from JSDoc comments.

Function name for LLM invocation.

The identifier used by the LLM to call this function. Must be unique within the application.

OpenAI limits function names to 64 characters.

Schema for the return type.

Defines the expected output type as an object parameters schema, wrapping the return type in an ILlmSchema.IParameters object with $defs for shared type definitions and properties for the structured output.

undefined when the function returns void or has no meaningful return value.

Schema for function parameters.

Defines the expected argument types as a JSON Schema object. Contains $defs for shared type definitions and properties for each named parameter.

Lenient JSON parser with schema-based coercion.

Handles incomplete/malformed JSON commonly produced by LLMs:

• Unclosed brackets, strings, trailing commas
• JavaScript-style comments (// and multi-line)
• Unquoted object keys, incomplete keywords (tru, fal, nul)
• Markdown code block extraction, junk prefix skipping

Also coerces double-stringified values ("42" → 42) using the parameters schema.

Type validation is NOT performed — use validate after parsing.

If the SDK (e.g., LangChain, Vercel AI, MCP) already parses JSON internally and provides a pre-parsed object, use coerce instead to apply schema-based type coercion without re-parsing.

Category tags for function organization.

Extracted from @tag JSDoc annotations. Useful for grouping related functions and filtering the function list.

Validates LLM-generated arguments against the schema.

LLMs frequently make type errors such as returning strings instead of numbers or missing required properties. Use this validator to check arguments before execution.

When validation fails, use LlmJson.stringify from @typia/utils to format the error for LLM feedback. The formatted output shows the invalid JSON with inline error comments, helping the LLM understand and correct its mistakes in the next turn.