Enable tool-use ability for any LLM model (DeepSeek V3/R1, etc.)
For some models/providers that doesn't natively support function calling (e.g. DeepSeek V3/R1), you can use this library to transform the tool calls to a user prompt, in Hermes template format by default.
pip install tooluserfrom openai import AsyncOpenAI
from tooluser import make_tool_user
oai = make_tool_user(AsyncOpenAI())
res = await oai.chat.completions.create(
model="deepseek/deepseek-chat-v3-0324", # From OpenRouter https://openrouter.ai/deepseek/deepseek-chat-v3-0324
messages=[{"role": "user", "content": "What's the time in Shanghai?"}],
tools=[
{
"type": "function",
"function": {
"name": "get_time",
"description": "Get the time in a given location",
"parameters": {
"type": "object",
"properties": {
"location": {
"type": "string",
"description": "The location to get the time for",
},
},
},
},
}
],
)Check out the example.py for a runnable example.
Yes, this library also supports streaming.
Check out the example_stream.py for a runnable example.
(LLM output for tool using is not streamed, because we use json-repair for it.)
Some LLMs occasionally forget to wrap function calls in <tool_call> tags and output raw JSON instead. This library can optionally detect such cases when they appear at the end of the response.
from tooluser import make_tool_user
# Enable raw JSON detection
client = make_tool_user(
AsyncOpenAI(),
enable_raw_json_detection=True
)Example scenarios that will be detected:
"I'll help you with that. {"name": "get_weather", "arguments": {"location": "NYC"}}""Let me search for that information. {"name": "search_files", "arguments": {"pattern": "*.py"}}"
What won't be detected (to avoid false positives):
"Here's some data: {"name": "config", "arguments": {...}} for processing"- JSON that appears in the middle of the response
Note: This feature is disabled by default for maximum reliability. Only enable it if you're experiencing issues with LLMs that inconsistently use tool call tags.
Check out the example_raw_json.py for a runnable example.
Function calling is implicitly a prompt template, to make the model understand how to output the structured response as we want. Hermes template is a widely adopted prompt template for function calling.
- Qwen2.5 - Function Calling Templates
- Qwen2.5-0.5B-Instruct - Tokenizer Config
- Hermes-Function-Calling
As we want to make use of the OpenAI chat completion API, we do not directly use Hermes template to generate the LLM instruction, but we generate the Hermes style system prompt and user prompt.
The actually API call is:
System:
<tool_instruction>
You are a function calling AI model. You are provided with function signatures within <tools> </tools> XML tags. You may call one or more functions to assist with the user query. Don't make assumptions about what values to plug into functions.
<tools>
['{"type": "function", "function": {"name": "get_time", "description": "Get the time in a given location", "parameters": {"type": "object", "properties": {"location": {"type": "string", "description": "The location to get the time for"}}}}}']
</tools>
For each function call return a json object with function name and arguments within <tool_call> </tool_call> tags with the following schema:
<tool_call>
{"name": <function-name>, "arguments": <args-dict>}
</tool_call>
Here is an example of a tool call:
<tool_call>
{"name": "get_weather", "arguments": {"location": "San Francisco, CA", "unit": "celsius"}}
</tool_call>
</tool_instruction>User:
What's the time in Shanghai?Assistant:
<tool_call>
{"name": "get_time", "arguments": {"location": "Shanghai"}}
</tool_call>