Hermes-Function-Calling

Hermes-Function-Calling is an open-source repository by NousResearch that enables the Hermes Pro Large Language Model to perform structured function calling and JSON mode inference. It provides a complete inference pipeline with recursive tool-call execution, Pydantic-based schema validation, and OpenAI-compatible ChatML prompt formatting. The project supports both function calling (where the model selects and invokes tools) and JSON mode (where the model generates structured outputs conforming to a given schema).

• Function Calling Inference — Run functioncall.py with a natural language query; the model selects the appropriate tool, generates a structured tool_call JSON, executes it, and returns a natural language response.
• JSON Mode / Structured Outputs — Use jsonmode.py with a Pydantic model schema to force the model to respond with a valid JSON object matching your schema.
• Custom Function Support — Add new tools by defining Python functions decorated with @tool in functions.py and registering them in get_openai_tools().
• Custom Pydantic Models — Define your own Pydantic BaseModel classes and serialize them to JSON schema for use in JSON mode inference.
• ChatML Prompt Format — Uses the ChatML format with <|im_start|> / <|im_end|> tokens, compatible with OpenAI API conventions and multi-turn dialogue.
• Hermes-3 GOAP Reasoning — Supports Goal Oriented Action Planning (GOAP) via <scratch_pad> tags for structured reasoning before tool invocation.
• 4-bit Quantization Support — Load models in 4-bit precision using bitsandbytes via the --load_in_4bit flag for reduced memory usage.
• Few-Shot Examples — Optionally include few-shot examples via --num_fewshot to guide model behavior.
• Recursive Depth Control — Configure maximum recursive function call iterations with --max_depth to prevent infinite loops.
• yfinance Integration — Built-in financial data functions (stock prices, fundamentals, financial statements) using the yfinance library as demonstration tools.