pymllm.parsers.tool_call_parser¶

Tool-call (function-calling) output parser.

Extracts structured tool calls from model output text. Supports both one-shot and incremental streaming modes.

Formats supported:

qwen25 — <tool_call>{"name":...,"arguments":...}</tool_call>
llama3 — <|python_tag|>{"name":...,"parameters":...}
hermes — <tool_call>{"name":...,"arguments":...}</tool_call> (same tags, Hermes schema)

Usage:

# Non-streaming
parser = ToolCallParser("qwen25", tools=tools_list)
content, tool_calls = parser.parse_non_stream(full_text)

# Streaming
parser = ToolCallParser("qwen25", tools=tools_list)
for delta in deltas:
    content_delta, tool_call_deltas = parser.parse_stream_chunk(delta)

Classes¶

`ToolCallItem`	A single parsed tool call.
`ToolCallParser`	Model-agnostic tool-call parser.

Module Contents¶

class pymllm.parsers.tool_call_parser.ToolCallItem¶

A single parsed tool call.

name: str | None = None¶

arguments: str = ''¶

tool_call_id: str = ''¶

index: int = 0¶

to_openai_dict(streaming=True)¶

Convert to OpenAI tool_calls[] element format.

Parameters:: streaming (bool) – If True, include index (streaming delta format). If False, omit index (non-streaming message format).
Return type:: Dict[str, Any]

class pymllm.parsers.tool_call_parser.ToolCallParser(model_type, tools=None)¶

Model-agnostic tool-call parser.

Parameters:

model_type (str) – Key into the format registry (e.g. "qwen25", "llama3").
tools (Optional[List[Any]]) – The tools list from the OpenAI chat request (used to resolve function names).

SUPPORTED¶

has_tool_call(text)¶

Return True if text contains a tool-call pattern.

Parameters:: text (str)
Return type:: bool

parse_non_stream(text)¶

Parse complete text.

Returns (remaining_content, tool_calls).

Parameters:: text (str)
Return type:: Tuple[str, List[ToolCallItem]]

parse_stream_chunk(delta)¶

Parse an incremental streaming delta.

Returns (content_delta, tool_call_items).

For tool call items: - First item for a call: name is set, arguments is "". - Subsequent items: name is None, arguments is the new

characters appended (argument delta).

Parameters:: delta (str)
Return type:: Tuple[str, List[ToolCallItem]]

flush()¶

Flush any remaining buffered tool call (call at request end).

Return type:: List[ToolCallItem]