Skip to content
Sonar

Reasoning and tool parsers

Reasoning and tool parsers convert model text into structured API response fields. The parser must match the format produced by the model checkpoint.

Configure reasoning

Terminal window
aphrodite serve MODEL --reasoning-parser glm45

The parser separates reasoning content from the final answer. Parser names are model-format names, not general quality settings.

List the registered values:

Terminal window
aphrodite serve --help=reasoning-parser

Test a known prompt and inspect both response fields before production use. Model template changes can alter delimiters that the parser expects.

Configure automatic tool choice

Terminal window
aphrodite serve MODEL \
  --enable-auto-tool-choice \
  --tool-call-parser glm47

Clients can now send tools and use tool_choice="auto". The model decides whether to call a tool, and the parser converts its output to the OpenAI tool call structure.

List current parser names:

Terminal window
aphrodite serve --help=tool-call-parser

Validate a model pair

Use the parser recommended by the checkpoint publisher. Check the chat template and parser together because the template controls the tool instructions.

Run these cases:

  • A request that should call one tool
  • A request that should answer without a tool
  • More than one tool in the request
  • Invalid tool arguments from the model
  • Streaming output

Applications must validate tool names and arguments. A parsed call is model output and remains untrusted.

Custom parsers

--reasoning-parser-plugin can load a custom reasoning parser. Plugin code runs inside the server process. Pin and review it as you would remote model code.

Troubleshoot parsing

Disable sampling with temperature zero. Capture the raw model output and compare its delimiters with the parser format.

If reasoning text appears in the final answer, check the reasoning parser. If a tool call remains plain text, check the tool parser and automatic tool-choice flag.