Concrete base class for chat format adapters (80% logic). More...

#include <entropic/inference/adapters/adapter_base.h>

Inheritance diagram for entropic::ChatAdapter:

[legend]

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

virtual ThinkMarkers	thinking_markers () const
	Parse tool calls from model output.

std::string	strip_think_blocks (const std::string &content) const
	Strip this family's reasoning blocks from content (gh#108).

virtual ParseResult	parse_tool_calls (const std::string &content) const =0

virtual Message	format_tool_result (const ToolCall &tool_call, const std::string &result) const
	Format a tool result as a user message.

virtual std::string	format_tools (const std::vector< std::string > &tool_jsons) const
	Format tool definitions for injection into system prompt.

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

virtual std::string	chat_format () const =0
	Chat format identifier (e.g.

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::vector< ToolCall >	parse_tagged_tool_calls (const std::string &content) const
	Parse <tool_call>JSON</tool_call> tagged blocks.

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.

Protected Attributes
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

Concrete base class for chat format adapters (80% logic).

Provides shared parsing primitives, think-block handling, JSON recovery, and system prompt assembly. Subclasses override parse_tool_calls() and optionally format_tools()/format_tool_result().

Version: 1.8.2

Definition at line 129 of file adapter_base.h.

Constructor & Destructor Documentation

◆ 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 103 of file adapter_base.cpp.

Member Function Documentation

◆ chat_format()

virtual std::string entropic::ChatAdapter::chat_format ( ) const

pure virtual

Chat format identifier (e.g.

"chatml").

Returns: Format string, or empty for GGUF-embedded template.

Version: 1.8.2

Implemented in entropic::Gemma4Adapter, entropic::GenericAdapter, entropic::Nemotron3Adapter, entropic::Qwen35Adapter, and entropic::Qwen36Adapter.

◆ format_content_parts()

std::string ChatAdapter::format_content_parts ( const std::vector< ContentPart > & parts ) const

virtual

Convert multimodal content parts to adapter-specific format.

Default: OpenAI-format content array JSON.

Parameters

parts Content parts from a message.

Returns: JSON string in the format expected by the model's chat template.

Version: 1.9.11

Default: OpenAI-format content array. Adapters override if the model expects a different image reference format.

Parameters

parts Content parts from a message.

Returns: JSON array string with text/image objects. @dg_internal

Version: 1.9.11

Reimplemented in entropic::Qwen35Adapter, and entropic::Qwen36Adapter.

Definition at line 604 of file adapter_base.cpp.

◆ format_system_with_vision()

std::string ChatAdapter::format_system_with_vision	(	const std::string &	base_system,
		bool	has_vision
	)		const

virtual

Format system prompt with optional vision context.

Default: return system prompt unchanged regardless of vision.

Parameters

base_system	Base system prompt text.
has_vision	Whether the model has vision capability.

Returns: Formatted system prompt.

Version: 1.9.11

Default implementation returns base_system unchanged. Vision-capable adapters override to append vision instructions.

Parameters

base_system	Base system prompt text.
has_vision	Whether vision is available.

Returns: base_system unchanged. @dg_internal

Version: 1.9.11

Reimplemented in entropic::Qwen35Adapter, and entropic::Qwen36Adapter.

Definition at line 591 of file adapter_base.cpp.

◆ format_tool_result()

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

virtual

Format a tool result as a user message.

Format tool result as user message (default).

Parameters

tool_call	The tool call that was executed.
result	Tool execution result text.

Returns: Formatted message.

Version: 1.8.2

Parameters

tool_call	The executed tool call.
result	Execution result text.

Returns: Formatted user message. @dg_internal

Version: 1.8.2

Reimplemented in entropic::Nemotron3Adapter, entropic::Qwen35Adapter, and entropic::Qwen36Adapter.

Definition at line 121 of file adapter_base.cpp.

◆ format_tools()

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

virtual

Format tool definitions for injection into system prompt.

Default tool formatting: markdown headings + JSON schema.

Parameters

tool_jsons Tool definition JSON strings.

Returns: Adapter-formatted tool prompt string.

Version: 2.0.4

Parameters

tool_jsons Tool definitions as JSON strings.

Returns: Formatted tool section string. @dg_internal

Version: 1.8.2

Reimplemented in entropic::Nemotron3Adapter, entropic::Qwen35Adapter, and entropic::Qwen36Adapter.

Definition at line 535 of file adapter_base.cpp.

◆ is_response_complete()

bool entropic::ChatAdapter::is_response_complete	(	const std::string &	content,
		const std::vector< ToolCall > &	tool_calls
	)		const

Check if response represents task completion.

Parameters

content	Response content.
tool_calls	Parsed tool calls.

Returns: true if complete.

Version: 1.8.2

Think-aware: think-only content without tool calls = still working.

Parameters

content	Response content.
tool_calls	Parsed tool calls.

Returns: true if complete. @dg_internal

Version: 1.8.2

Definition at line 145 of file adapter_base.cpp.

◆ parse_single_tool_call()

std::optional< ToolCall > ChatAdapter::parse_single_tool_call ( const std::string & json_str ) const

protected

Parse a single JSON tool call string.

Parameters

json_str JSON from tagged block.

Returns: Parsed ToolCall or nullopt.

Version: 1.8.2

Parameters

json_str JSON string from tagged block.

Returns: Parsed ToolCall or nullopt. @dg_internal

Version: 2.3.8

Definition at line 567 of file adapter_base.cpp.

◆ parse_tagged_tool_calls()

std::vector< ToolCall > entropic::ChatAdapter::parse_tagged_tool_calls ( const std::string & content ) const

protected

Parse <tool_call>JSON</tool_call> tagged blocks.

Version: 1.8.2

Part of the ~80% the concrete base owns; concrete adapters override only their family specifics. Tolerates the asymmetric open tags real families emit (gemma4's pipe-prefixed <|tool_call> against a plain close).

Parameters

content Model output content.

Returns: Vector of parsed tool calls; empty when no tagged block parses. @req REQ-INFER-012

Version: 2.3.8

Definition at line 180 of file adapter_base.cpp.

◆ parse_tool_calls()

virtual ParseResult entropic::ChatAdapter::parse_tool_calls ( const std::string & content ) const

pure virtual

Implemented in entropic::Gemma4Adapter, entropic::GenericAdapter, entropic::Nemotron3Adapter, entropic::Qwen35Adapter, and entropic::Qwen36Adapter.

◆ strip_think_blocks()

std::string entropic::ChatAdapter::strip_think_blocks ( const std::string & content ) const

Strip this family's reasoning blocks from content (gh#108).

Uses thinking_markers(), so an adapter declares its markers once and both the buffered strip and the live StreamThinkFilter agree. Matching is plain substring, not regex — <|channel> contains a regex metacharacter and would need escaping.

An unclosed opening marker (generation hit its token budget mid-reasoning) erases to end of content and WARNs: no answer was ever produced, so returning the reasoning would be worse than nothing.

PUBLIC since v2.10.3 — was protected. The shared parse rule (response_parse.h) runs it on both the template and adapter branches, so it is an inference-layer concern now, not a subclass helper. Widening only; non-virtual, so no ABI change for existing consumers.

Parameters

content Model output.

Returns: Content with reasoning blocks removed and trimmed.

Version: 2.10.3

Driven by thinking_markers() so an adapter declares its delimiters once. Plain substring matching, not regex: gemma4's <|channel> contains a regex metacharacter and the previous std::regex form would have needed escaping to stay correct.

An opening marker with no close means the budget ran out mid-reasoning: the span is erased to end of content and a WARN is logged, because no answer was ever produced and returning the reasoning would be worse than nothing.

Parameters

content Model output.

Returns: Content with this family's reasoning blocks removed and trimmed; unchanged when the family declares no markers. @req REQ-INFER-011

Version: 2.11.0

Definition at line 426 of file adapter_base.cpp.

◆ thinking_markers()

virtual ThinkMarkers entropic::ChatAdapter::thinking_markers ( ) const

inlinevirtual

Parse tool calls from model output.

Parameters

content Raw model output.

Returns: ParseResult with cleaned content and tool calls.

Version: 1.8.2

This family's reasoning-block delimiters (gh#108).

Default is the <think> pair used by Qwen and Nemotron. Override in a family that wraps reasoning differently — Gemma-4 QAT does.

The single declaration point per family: both consumers read it — the buffered strip (strip_think_blocks) and the live StreamThinkFilter that wraps on_token in generate_streaming. Before v2.10.3 these were two independent hardcoded literals and gemma4 matched neither.

Returns: Marker pair; empty for a family that declares none, which makes both consumers no-ops. @req REQ-INFER-011

Version: 2.10.3

Reimplemented in entropic::Gemma4Adapter.

Definition at line 167 of file adapter_base.h.

◆ try_recover_json()

std::optional< ToolCall > entropic::ChatAdapter::try_recover_json ( const std::string & json_str ) const

protected

Attempt JSON recovery on malformed tool call string.

Version: 1.8.2

Member Data Documentation

◆ identity_prompt_

std::string entropic::ChatAdapter::identity_prompt_

protected

Assembled identity prompt.

Definition at line 269 of file adapter_base.h.

◆ tier_name_

std::string entropic::ChatAdapter::tier_name_

protected

Identity tier name.

Definition at line 268 of file adapter_base.h.

◆ tool_prefixes_

std::unordered_set<std::string> entropic::ChatAdapter::tool_prefixes_

mutableprotected

Known tool prefixes.

Definition at line 270 of file adapter_base.h.

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

entropic/inference/adapters/adapter_base.h
inference/adapters/adapter_base.cpp

Public Member Functions

Protected Member Functions

Protected Attributes

Detailed Description

Constructor & Destructor Documentation

◆ ChatAdapter()

Member Function Documentation

◆ chat_format()

◆ format_content_parts()

◆ format_system_with_vision()

◆ format_tool_result()

◆ format_tools()

◆ is_response_complete()

◆ parse_single_tool_call()

◆ parse_tagged_tool_calls()

◆ parse_tool_calls()

◆ strip_think_blocks()

◆ thinking_markers()

◆ try_recover_json()

Member Data Documentation

◆ identity_prompt_

◆ tier_name_

◆ tool_prefixes_