sync update from LAN repo (source ji @ 85078a31)

This commit is contained in:
gozeros committed 2026-10-21 15:09:16 +08:00
1 parent 5d6627af50
commit e602d88695
864 files changed
+99150 -6346

No files matched your search

@@ -19,8 +19,9 @@ re-inject thinking metadata back into the outgoing payload for any assistant
message that originally carried it:
- ``thought_signature`` for Gemini-compatible tool calls.
- ``reasoning_content`` for providers (for example DeepSeek) that require it
to be echoed in multi-turn requests when thinking mode is enabled.
- ``reasoning_content`` and the newer ``reasoning`` spelling used by vLLM and
other OpenAI-compatible providers. Both are normalized to
``additional_kwargs.reasoning_content`` for DeerFlow's existing clients.
"""
from __future__ import annotations
@@ -166,7 +167,7 @@ class PatchedChatOpenAI(ChatOpenAI):
default_chunk_class: type,
base_generation_info: dict | None,
) -> ChatGenerationChunk | None:
"""Preserve streamed ``reasoning_content`` in ``additional_kwargs``."""
"""Preserve streamed ``reasoning_content``/``reasoning`` metadata."""
# Some compatible gateways return a vendor error event followed by a
# normal-looking usage chunk, all under HTTP 200. Detect the first
# event before LangChain converts it into an empty AIMessageChunk.
@@ -174,8 +175,8 @@ class PatchedChatOpenAI(ChatOpenAI):
generation_chunk = super()._convert_chunk_to_generation_chunk(
chunk,
default_chunk_class,
base_generation_info,
)
base_generation_info,
)
if generation_chunk is None:
return None
@@ -183,11 +184,12 @@ class PatchedChatOpenAI(ChatOpenAI):
choice = choices[0] if choices else {}
delta = choice.get("delta") if isinstance(choice, dict) else None
reasoning_content = _extract_reasoning_content(delta if isinstance(delta, dict) else None)
if reasoning_content and isinstance(generation_chunk.message, AIMessageChunk):
generation_chunk.message = _append_reasoning_content(
generation_chunk.message,
reasoning_content,
)
if reasoning_content and isinstance(generation_chunk.message, AIMessageChunk):
generation_chunk.message = _append_reasoning_content(
generation_chunk.message,
reasoning_content,
reasoning=delta.get("reasoning") if isinstance(delta, dict) else None,
)
return generation_chunk
def _create_chat_result(
@@ -195,20 +197,24 @@ class PatchedChatOpenAI(ChatOpenAI):
response: dict | Any,
generation_info: dict | None = None,
) -> ChatResult:
"""Preserve non-stream ``reasoning_content`` in ``additional_kwargs``."""
"""Preserve non-stream ``reasoning_content``/``reasoning`` metadata."""
response_dict = response if isinstance(response, dict) else response.model_dump()
_raise_for_application_error(response_dict)
result = super()._create_chat_result(response, generation_info)
choices = response_dict.get("choices", [])
generations: list[ChatGeneration] = []
generations: list[ChatGeneration] = []
for index, generation in enumerate(result.generations):
message = generation.message
choice = choices[index] if index < len(choices) else {}
choice_message = choice.get("message", {}) if isinstance(choice, dict) else {}
reasoning_content = _extract_reasoning_content(choice_message if isinstance(choice_message, dict) else None)
if reasoning_content and isinstance(message, AIMessage):
message = _append_reasoning_content(message, reasoning_content)
if reasoning_content and isinstance(message, AIMessage):
message = _append_reasoning_content(
message,
reasoning_content,
reasoning=choice_message.get("reasoning") if isinstance(choice_message, dict) else None,
)
generation = ChatGeneration(
message=message,
generation_info=generation.generation_info,
@@ -259,34 +265,46 @@ def _restore_tool_call_signatures(payload_msg: dict, orig_msg: AIMessage) -> Non
payload_tc["thought_signature"] = sig
def _restore_reasoning_content(payload_msg: dict, orig_msg: AIMessage) -> None:
"""Re-inject ``reasoning_content`` onto assistant payload message.
Some thinking-enabled OpenAI-compatible providers require each historical
assistant turn to include the original ``reasoning_content`` in subsequent
requests. LangChain stores it in ``AIMessage.additional_kwargs`` but does
not always serialize it back into ``messages``.
"""
reasoning_content = orig_msg.additional_kwargs.get("reasoning_content")
if reasoning_content is not None:
payload_msg["reasoning_content"] = reasoning_content
def _extract_reasoning_content(payload: dict | None) -> str | None:
"""Extract reasoning text from provider payload delta/message objects."""
if not isinstance(payload, dict):
return None
reasoning_content = payload.get("reasoning_content")
return reasoning_content if isinstance(reasoning_content, str) and reasoning_content else None
def _append_reasoning_content(message: AIMessage | AIMessageChunk, reasoning_content: str) -> AIMessage | AIMessageChunk:
"""Append streamed reasoning chunks, dedupe complete response reasoning."""
additional_kwargs = dict(message.additional_kwargs)
existing = additional_kwargs.get("reasoning_content")
def _restore_reasoning_content(payload_msg: dict, orig_msg: AIMessage) -> None:
"""Re-inject provider reasoning onto an assistant history message.
Some thinking-enabled OpenAI-compatible providers require each historical
assistant turn to include its original reasoning in subsequent requests.
Preserve the spelling returned by the provider when available; otherwise
retain the established ``reasoning_content`` behavior.
"""
reasoning = orig_msg.additional_kwargs.get("reasoning")
if reasoning is not None:
payload_msg["reasoning"] = reasoning
return
reasoning_content = orig_msg.additional_kwargs.get("reasoning_content")
if reasoning_content is not None:
payload_msg["reasoning_content"] = reasoning_content
def _extract_reasoning_content(payload: dict | None) -> str | None:
"""Extract either supported reasoning spelling from provider payloads."""
if not isinstance(payload, dict):
return None
reasoning_content = payload.get("reasoning_content") or payload.get("reasoning")
return reasoning_content if isinstance(reasoning_content, str) and reasoning_content else None
def _append_reasoning_content(
message: AIMessage | AIMessageChunk,
reasoning_content: str,
*,
reasoning: Any = None,
) -> AIMessage | AIMessageChunk:
"""Normalize reasoning for clients while retaining its provider spelling."""
additional_kwargs = dict(message.additional_kwargs)
existing = additional_kwargs.get("reasoning_content")
if isinstance(existing, str) and existing:
if reasoning_content not in existing:
additional_kwargs["reasoning_content"] = f"{existing}{reasoning_content}"
else:
additional_kwargs["reasoning_content"] = reasoning_content
return message.model_copy(update={"additional_kwargs": additional_kwargs})
else:
additional_kwargs["reasoning_content"] = reasoning_content
if reasoning is not None:
additional_kwargs["reasoning"] = reasoning
return message.model_copy(update={"additional_kwargs": additional_kwargs})