Functionary-small-v2.2: Advanced Function Calling Model
Functionary-small-v2.2, developed by MeetKai, is a 7 billion parameter language model engineered for sophisticated function calling and tool interaction. Unlike general-purpose LLMs, this model is specialized in understanding when to invoke functions, executing them in parallel or serially, and processing their outputs to generate grounded responses. It can also intelligently decide when not to use tools, providing standard chat responses instead.
Key Capabilities
- Intelligent Parallel Tool Use: Capable of orchestrating multiple tool calls concurrently.
- Output Grounding: Analyzes function/tool outputs to formulate relevant and accurate responses.
- Discretionary Tool Use: Determines the necessity of tool invocation, opting for direct chat responses when appropriate.
- OpenAI-like Function Calls: Utilizes JSON Schema Objects for function definitions, similar to OpenAI's GPT function calling mechanism.
Performance
On MeetKai's in-house dataset, Functionary-small-v2.2 achieves a Function Calling Accuracy (Name & Arguments) of 0.546, outperforming OpenAI's gpt-3.5-turbo-1106 (0.531) in this specific metric. This highlights its strong specialization in tool use scenarios.
Prompt Template
The model employs a unique "v2PromptTemplate" that structures turns into from, recipient, and content portions. Function definitions are converted into TypeScript-like definitions and injected as system prompts, followed by a default system prompt and then the conversation messages. This formatting is supported by MeetKai's vLLM server, which can automatically apply the chat template.
Good for
- Applications requiring precise and reliable function calling.
- Building agents that interact with external APIs and tools.
- Developers seeking an open-source model optimized for tool use, offering an alternative to larger proprietary models for specific tasks.