Nemotron 3 chat adapter (hybrid Mamba-Transformer family). More...

#include </home/runner/work/entropic/entropic/src/inference/adapters/nemotron3_adapter.h>

Inheritance diagram for entropic::Nemotron3Adapter:

Collaboration diagram for entropic::Nemotron3Adapter:

Public Member Functions
std::string	chat_format () const override
	Chat format: GGUF-embedded template (Nemotron-specific).

ParseResult	parse_tool_calls (const std::string &content) const override
	Parse DSML invoke calls; fall back to qwen XML, then tagged JSON.

Message	format_tool_result (const ToolCall &tool_call, const std::string &result) const override
	Wrap tool result in `<tool_response>` tags.

	ChatAdapter (std::string tier_name, std::string identity_prompt)
	Construct adapter with tier identity.

Public Member Functions inherited from entropic::ChatAdapter
	ChatAdapter (std::string tier_name, std::string identity_prompt)
	Construct adapter with tier identity.

std::string	format_system_prompt (const std::string &base_prompt, const std::vector< std::string > &tool_jsons) const
	Assemble system prompt: identity + context + tools.

bool	is_response_complete (const std::string &content, const std::vector< ToolCall > &tool_calls) const
	Check if response represents task completion.

virtual std::string	format_system_with_vision (const std::string &base_system, bool has_vision) const
	Format system prompt with optional vision context.

virtual std::string	format_content_parts (const std::vector< ContentPart > &parts) const
	Convert multimodal content parts to adapter-specific format.

Protected Member Functions
std::string	format_tools (const std::vector< std::string > &tool_jsons) const override
	Format tools as a `<tools>` JSON array, then teach DSML invoke.

Protected Member Functions inherited from entropic::ChatAdapter
std::vector< ToolCall >	parse_tagged_tool_calls (const std::string &content) const
	Parse <tool_call>JSON</tool_call> tagged blocks.

std::vector< ToolCall >	parse_bare_json_tool_calls (const std::string &content) const
	Parse bare JSON lines containing "name" key.

std::string	extract_thinking (const std::string &content) const
	Extract <think>...</think> content.

std::string	strip_think_blocks (const std::string &content) const
	Strip all <think>...</think> blocks from content.

std::optional< ToolCall >	try_recover_json (const std::string &json_str) const
	Attempt JSON recovery on malformed tool call string.

std::optional< ToolCall >	parse_single_tool_call (const std::string &json_str) const
	Parse a single JSON tool call string.

Additional Inherited Members
Protected Attributes inherited from entropic::ChatAdapter
std::string	tier_name_
	Identity tier name.

std::string	identity_prompt_
	Assembled identity prompt.

std::unordered_set< std::string >	tool_prefixes_
	Known tool prefixes.

Detailed Description

Nemotron 3 chat adapter (hybrid Mamba-Transformer family).

Tool-call parsing mirrors the qwen3_coder XML format. Reasoning traces are emitted as <think>...</think> blocks and handled by the shared base-class primitives.

Version: 2.1.9

Definition at line 60 of file nemotron3_adapter.h.

Member Function Documentation

◆ chat_format()

std::string entropic::Nemotron3Adapter::chat_format ( ) const

inlineoverridevirtual

Chat format: GGUF-embedded template (Nemotron-specific).

Returns: Empty string — llama.cpp drives template application from the GGUF's stored template. @utility

Version: 2.1.9

Implements entropic::ChatAdapter.

Definition at line 71 of file nemotron3_adapter.h.

◆ ChatAdapter()

entropic::ChatAdapter::ChatAdapter	(	std::string	tier_name,
		std::string	identity_prompt
	)

Construct adapter with tier identity.

Parameters

tier_name	Identity tier (e.g. "eng", "lead").
identity_prompt	Assembled identity prompt.

Version: 1.8.2

Parameters

tier_name	Identity tier name.
identity_prompt	Assembled identity prompt.

Version: 1.8.2

Definition at line 66 of file adapter_base.cpp.

◆ format_tool_result()

Message entropic::Nemotron3Adapter::format_tool_result	(	const ToolCall &	tool_call,
		const std::string &	result
	)		const

overridevirtual

Wrap tool result in <tool_response> tags.

Wrap tool result in <tool_response> tags as a user message.

Parameters

tool_call	Executed tool call.
result	Execution result text.

Returns: User-role message containing the wrapped result.

Version: 2.1.9

Parameters

tool_call	Executed tool call (used for logging only).
result	Execution result text.

Returns: User-role message with the wrapped result.

Reimplemented from entropic::ChatAdapter.

Definition at line 235 of file nemotron3_adapter.cpp.

◆ format_tools()

std::string entropic::Nemotron3Adapter::format_tools ( const std::vector< std::string > & tool_jsons ) const

overrideprotectedvirtual

Format tools as a <tools> JSON array, then teach DSML invoke.

Format tool definitions and teach the DSML invoke call format.

Parameters

tool_jsons Tool definition JSON strings.

Returns: Section to inject into the system prompt.

Version: 2.3.8

gh#70: the injected call-format example matches what the nemotron_h GGUFs actually emit (DSML invoke, fullwidth-pipe ｜), instead of the qwen XML the model never produced under the production prompt.

Parameters

tool_jsons Tool definition JSON strings.

Returns: Section to inject into the system prompt.

Reimplemented from entropic::ChatAdapter.

Definition at line 290 of file nemotron3_adapter.cpp.

◆ parse_tool_calls()

ParseResult entropic::Nemotron3Adapter::parse_tool_calls ( const std::string & content ) const

overridevirtual

Parse DSML invoke calls; fall back to qwen XML, then tagged JSON.

Parse tool calls from Nemotron 3 output.

Parameters

content Raw model output.

Returns: ParseResult.

Version: 2.3.8

Layered strategy (gh#70): DSML invoke first (the native emit), then the qwen3_coder XML path and the tagged-JSON path as backstops for rigged-prompt or mixed-format consumers.

Parameters

content Raw model output.

Returns: ParseResult.

Implements entropic::ChatAdapter.

Definition at line 60 of file nemotron3_adapter.cpp.

The documentation for this class was generated from the following files:

inference/adapters/nemotron3_adapter.h
inference/adapters/nemotron3_adapter.cpp

Public Member Functions

Protected Member Functions

Additional Inherited Members

Detailed Description

Member Function Documentation

◆ chat_format()

◆ ChatAdapter()

◆ format_tool_result()

◆ format_tools()

◆ parse_tool_calls()