Parsers

The parsers module provides utilities for parsing tool calls from model outputs and formatting chat templates.

ToolParser

Abstract base class for tool call parsers.

from rllm.parser import ToolParser

Methods

parse

Extract tool calls from model response.

tool_calls = parser.parse(model_response)

str

Raw model output containing tool calls.

list[ToolCall]

List of parsed tool calls.

get_tool_prompt

Get the tool prompt for the model.

prompt = parser.get_tool_prompt(tools_schema)

str

JSON schema for available tools.

str

Formatted tool prompt to include in system message.

get_parser

Factory method to get appropriate parser for a tokenizer.

parser = ToolParser.get_parser(tokenizer)

PreTrainedTokenizer

Tokenizer to detect parser type from.

ToolParser

Appropriate parser instance (R1ToolParser or QwenToolParser).

R1ToolParser

Parser for DeepSeek R1-style tool call format.

from rllm.parser import R1ToolParser

parser = R1ToolParser()

Format

R1 format uses special tokens:

<|tool_calls_begin|>
<|tool_call_begin|>function<|tool_sep|>function_name
{"param1": "value1", "param2": "value2"}
<|tool_call_end|>
<|tool_calls_end|>

Methods

tool_calls = parser.parse(model_response)
prompt = parser.get_tool_prompt(tools_schema)

QwenToolParser

Parser for Qwen-style tool call format.

from rllm.parser import QwenToolParser

parser = QwenToolParser()

Format

Qwen format uses XML-like tags:

<tool_call>{"name": "function_name", "arguments": {...}}</tool_call>

Methods

tool_calls = parser.parse(model_response)
prompt = parser.get_tool_prompt(tools_schema)

ChatTemplateParser

Parser for converting between chat format and tokens.

from rllm.parser import ChatTemplateParser

parser = ChatTemplateParser.get_parser(tokenizer)

Methods

apply_chat_template

Convert messages to tokens.

tokens = parser.apply_chat_template(
    messages,
    add_generation_prompt=True
)

decode

Convert tokens back to text.

text = parser.decode(tokens)

Example: Parsing Tool Calls (R1)

from rllm.parser import R1ToolParser
from rllm.tools import ToolCall

parser = R1ToolParser()

model_response = """
<|tool_calls_begin|>
<|tool_call_begin|>function<|tool_sep|>search_web
{"query": "What is rLLM?", "num_results": 5}
<|tool_call_end|>
<|tool_calls_end|>
"""

tool_calls = parser.parse(model_response)

for call in tool_calls:
    print(f"Tool: {call.name}")
    print(f"Arguments: {call.arguments}")

# Output:
# Tool: search_web
# Arguments: {'query': 'What is rLLM?', 'num_results': 5}

Example: Parsing Tool Calls (Qwen)

from rllm.parser import QwenToolParser

parser = QwenToolParser()

model_response = '''
Let me search for that information.
<tool_call>{"name": "search", "arguments": {"query": "capital of France"}}</tool_call>
'''

tool_calls = parser.parse(model_response)

for call in tool_calls:
    print(f"Tool: {call.name}")
    print(f"Arguments: {call.arguments}")

# Output:
# Tool: search
# Arguments: {'query': 'capital of France'}

Example: Multiple Tool Calls

from rllm.parser import QwenToolParser

parser = QwenToolParser()

model_response = '''
<tool_call>{"name": "calculator", "arguments": {"a": 5, "b": 3, "op": "add"}}</tool_call>
<tool_call>{"name": "calculator", "arguments": {"a": 10, "b": 2, "op": "multiply"}}</tool_call>
'''

tool_calls = parser.parse(model_response)

print(f"Found {len(tool_calls)} tool calls")
for i, call in enumerate(tool_calls):
    print(f"\nCall {i+1}:")
    print(f"  Name: {call.name}")
    print(f"  Args: {call.arguments}")

Example: Getting Tool Prompt

from rllm.parser import R1ToolParser
import json

parser = R1ToolParser()

tools_schema = json.dumps([
    {
        "type": "function",
        "function": {
            "name": "search_web",
            "description": "Search the web for information",
            "parameters": {
                "type": "object",
                "properties": {
                    "query": {"type": "string", "description": "Search query"}
                },
                "required": ["query"]
            }
        }
    }
], indent=2)

tool_prompt = parser.get_tool_prompt(tools_schema)

print(tool_prompt)
# Outputs formatted tool prompt with instructions

Example: Auto-detecting Parser

from transformers import AutoTokenizer
from rllm.parser import ToolParser

# DeepSeek model -> R1ToolParser
tokenizer = AutoTokenizer.from_pretrained("deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B")
parser = ToolParser.get_parser(tokenizer)
print(type(parser).__name__)  # R1ToolParser

# Qwen model -> QwenToolParser
tokenizer = AutoTokenizer.from_pretrained("Qwen/Qwen3-4B")
parser = ToolParser.get_parser(tokenizer)
print(type(parser).__name__)  # QwenToolParser

Example: Parsing Tool Calls

from transformers import AutoTokenizer
from rllm.parser import ToolParser

tokenizer = AutoTokenizer.from_pretrained("Qwen/Qwen3-4B")
parser = ToolParser.get_parser(tokenizer)

# Parse a model response into structured tool calls.
response = """<tool_call>
{"name": "search", "arguments": {"query": "rllm"}}
</tool_call>"""

tool_calls = parser.parse(response)
for tc in tool_calls:
    print(tc.name, tc.arguments)

Example: ChatTemplateParser

from transformers import AutoTokenizer
from rllm.parser import ChatTemplateParser

tokenizer = AutoTokenizer.from_pretrained("deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B")
parser = ChatTemplateParser.get_parser(tokenizer)

messages = [
    {"role": "system", "content": "You are a helpful assistant."},
    {"role": "user", "content": "What is 2+2?"}
]

# Convert to tokens
tokens = parser.apply_chat_template(
    messages,
    add_generation_prompt=True,
    return_tensors="pt"
)

print(f"Tokens: {tokens}")

# Decode back to text
text = parser.decode(tokens[0])
print(f"Formatted: {text}")

Core API

Engines

Training

Tools & utilities

ToolParser

Methods

parse

get_tool_prompt

get_parser

R1ToolParser

Format

Methods

QwenToolParser

Format

Methods

ChatTemplateParser

Methods

apply_chat_template

decode

Example: Parsing Tool Calls (R1)

Example: Parsing Tool Calls (Qwen)

Example: Multiple Tool Calls

Example: Getting Tool Prompt

Example: Auto-detecting Parser

Example: Parsing Tool Calls

Example: ChatTemplateParser

​ToolParser

​Methods

​parse

​get_tool_prompt

​get_parser

​R1ToolParser

​Format

​Methods

​QwenToolParser

​Format

​Methods

​ChatTemplateParser

​Methods

​apply_chat_template

​decode

​Example: Parsing Tool Calls (R1)

​Example: Parsing Tool Calls (Qwen)

​Example: Multiple Tool Calls

​Example: Getting Tool Prompt

​Example: Auto-detecting Parser

​Example: Parsing Tool Calls

​Example: ChatTemplateParser

ToolParser

Methods

parse

get_tool_prompt

get_parser

R1ToolParser

Format

Methods

QwenToolParser

Format

Methods

ChatTemplateParser

Methods

apply_chat_template

decode

Example: Parsing Tool Calls (R1)

Example: Parsing Tool Calls (Qwen)

Example: Multiple Tool Calls

Example: Getting Tool Prompt

Example: Auto-detecting Parser

Example: Parsing Tool Calls

Example: ChatTemplateParser