fix: normalize system messages before model calls
This commit is contained in:
1 parent
a28ad2140e
commit
94cf37a2ae
9 files changed
+236
-52
No files matched your search
@@ -0,0 +1,82 @@
|
||||
"""Utilities for provider-specific chat message ordering constraints."""
|
||||
|
||||
from __future__ import annotations
|
||||
|
||||
from typing import Any
|
||||
|
||||
from langchain_core.messages import BaseMessage, SystemMessage
|
||||
|
||||
|
||||
_SYSTEM_SEPARATOR = "\n\n---\n\n"
|
||||
|
||||
|
||||
def normalize_chat_payload_messages(payload: dict[str, Any]) -> None:
|
||||
"""Ensure payload ``system`` messages are represented once at the beginning.
|
||||
|
||||
Several OpenAI-compatible gateways reject requests when a transient
|
||||
``system`` message appears after user/assistant/tool history. LangChain and
|
||||
DeerFlow middleware can legitimately create those transient messages, so the
|
||||
provider boundary folds every system payload into a single leading message
|
||||
while preserving all non-system messages in their original order.
|
||||
"""
|
||||
messages = payload.get("messages")
|
||||
if not isinstance(messages, list):
|
||||
return
|
||||
|
||||
normalized = _normalize_payload_message_list(messages)
|
||||
if normalized is not messages:
|
||||
payload["messages"] = normalized
|
||||
|
||||
|
||||
def normalize_langchain_messages(messages: list[BaseMessage]) -> list[BaseMessage]:
|
||||
"""Return LangChain messages with all system context merged at the front."""
|
||||
system_messages = [message for message in messages if isinstance(message, SystemMessage)]
|
||||
if not system_messages:
|
||||
return messages
|
||||
|
||||
first_non_system_index = next((idx for idx, message in enumerate(messages) if not isinstance(message, SystemMessage)), len(messages))
|
||||
if len(system_messages) == first_non_system_index:
|
||||
return messages
|
||||
|
||||
merged = system_messages[0].model_copy(
|
||||
update={
|
||||
"content": _merge_contents([message.content for message in system_messages]),
|
||||
}
|
||||
)
|
||||
return [merged, *(message for message in messages if not isinstance(message, SystemMessage))]
|
||||
|
||||
|
||||
def _normalize_payload_message_list(messages: list[Any]) -> list[Any]:
|
||||
system_messages = [message for message in messages if isinstance(message, dict) and message.get("role") == "system"]
|
||||
if not system_messages:
|
||||
return messages
|
||||
|
||||
first_non_system_index = next(
|
||||
(
|
||||
idx
|
||||
for idx, message in enumerate(messages)
|
||||
if not (isinstance(message, dict) and message.get("role") == "system")
|
||||
),
|
||||
len(messages),
|
||||
)
|
||||
if len(system_messages) == first_non_system_index:
|
||||
return messages
|
||||
|
||||
merged_system = dict(system_messages[0])
|
||||
merged_system["content"] = _merge_contents([message.get("content") for message in system_messages])
|
||||
return [merged_system, *(message for message in messages if not (isinstance(message, dict) and message.get("role") == "system"))]
|
||||
|
||||
|
||||
def _merge_contents(contents: list[Any]) -> Any:
|
||||
if all(isinstance(content, str) for content in contents):
|
||||
return _SYSTEM_SEPARATOR.join(content for content in contents if content)
|
||||
|
||||
blocks: list[Any] = []
|
||||
for idx, content in enumerate(contents):
|
||||
if idx > 0:
|
||||
blocks.append({"type": "text", "text": _SYSTEM_SEPARATOR})
|
||||
if isinstance(content, list):
|
||||
blocks.extend(content)
|
||||
elif content is not None:
|
||||
blocks.append({"type": "text", "text": str(content)})
|
||||
return blocks
|
||||
+7
-5
@@ -6,10 +6,12 @@ import uuid
|
||||
from collections.abc import Iterator
|
||||
|
||||
import httpx
|
||||
from langchain_core.messages import AIMessage, AIMessageChunk, HumanMessage, ToolMessage
|
||||
from langchain_core.outputs import ChatGenerationChunk, ChatResult
|
||||
from langchain_openai import ChatOpenAI
|
||||
|
||||
from langchain_core.messages import AIMessage, AIMessageChunk, HumanMessage, ToolMessage
|
||||
from langchain_core.outputs import ChatGenerationChunk, ChatResult
|
||||
from langchain_openai import ChatOpenAI
|
||||
|
||||
from deerflow.models.message_ordering import normalize_langchain_messages
|
||||
|
||||
|
||||
def _fix_messages(messages: list) -> list:
|
||||
"""Sanitize incoming messages for MindIE compatibility.
|
||||
@@ -55,7 +57,7 @@ def _fix_messages(messages: list) -> list:
|
||||
|
||||
fixed.append(msg.model_copy(update={"content": text}))
|
||||
|
||||
return fixed
|
||||
return normalize_langchain_messages(fixed)
|
||||
|
||||
|
||||
def _parse_xml_tool_call_to_dict(content: str) -> tuple[str, list[dict]]:
|
||||
|
||||
+13
-10
@@ -9,10 +9,12 @@ on all assistant messages when thinking mode is enabled.
|
||||
|
||||
from typing import Any
|
||||
|
||||
from langchain_core.language_models import LanguageModelInput
|
||||
from langchain_core.messages import AIMessage
|
||||
from langchain_deepseek import ChatDeepSeek
|
||||
|
||||
from langchain_core.language_models import LanguageModelInput
|
||||
from langchain_core.messages import AIMessage
|
||||
from langchain_deepseek import ChatDeepSeek
|
||||
|
||||
from deerflow.models.message_ordering import normalize_chat_payload_messages
|
||||
|
||||
|
||||
class PatchedChatDeepSeek(ChatDeepSeek):
|
||||
"""ChatDeepSeek with proper reasoning_content preservation.
|
||||
@@ -65,9 +67,10 @@ class PatchedChatDeepSeek(ChatDeepSeek):
|
||||
ai_messages = [m for m in original_messages if isinstance(m, AIMessage)]
|
||||
assistant_payloads = [(i, m) for i, m in enumerate(payload_messages) if m.get("role") == "assistant"]
|
||||
|
||||
for (idx, payload_msg), ai_msg in zip(assistant_payloads, ai_messages):
|
||||
reasoning_content = ai_msg.additional_kwargs.get("reasoning_content")
|
||||
if reasoning_content is not None:
|
||||
payload_messages[idx]["reasoning_content"] = reasoning_content
|
||||
|
||||
return payload
|
||||
for (idx, payload_msg), ai_msg in zip(assistant_payloads, ai_messages):
|
||||
reasoning_content = ai_msg.additional_kwargs.get("reasoning_content")
|
||||
if reasoning_content is not None:
|
||||
payload_messages[idx]["reasoning_content"] = reasoning_content
|
||||
|
||||
normalize_chat_payload_messages(payload)
|
||||
return payload
|
||||
+11
-8
@@ -20,11 +20,13 @@ from langchain_core.language_models import LanguageModelInput
|
||||
from langchain_core.messages import AIMessage, AIMessageChunk
|
||||
from langchain_core.outputs import ChatGeneration, ChatGenerationChunk, ChatResult
|
||||
from langchain_openai import ChatOpenAI
|
||||
from langchain_openai.chat_models.base import (
|
||||
_convert_delta_to_message_chunk,
|
||||
_create_usage_metadata,
|
||||
)
|
||||
|
||||
from langchain_openai.chat_models.base import (
|
||||
_convert_delta_to_message_chunk,
|
||||
_create_usage_metadata,
|
||||
)
|
||||
|
||||
from deerflow.models.message_ordering import normalize_chat_payload_messages
|
||||
|
||||
_THINK_TAG_RE = re.compile(r"<think>\s*(.*?)\s*</think>", re.DOTALL)
|
||||
|
||||
|
||||
@@ -112,9 +114,10 @@ class PatchedChatMiniMax(ChatOpenAI):
|
||||
**extra_body,
|
||||
"reasoning_split": True,
|
||||
}
|
||||
else:
|
||||
payload["extra_body"] = {"reasoning_split": True}
|
||||
return payload
|
||||
else:
|
||||
payload["extra_body"] = {"reasoning_split": True}
|
||||
normalize_chat_payload_messages(payload)
|
||||
return payload
|
||||
|
||||
def _convert_chunk_to_generation_chunk(
|
||||
self,
|
||||
|
||||
+15
-12
@@ -29,9 +29,11 @@ from typing import Any
|
||||
|
||||
from langchain_core.language_models import LanguageModelInput
|
||||
from langchain_core.messages import AIMessage, AIMessageChunk
|
||||
from langchain_core.outputs import ChatGeneration, ChatGenerationChunk, ChatResult
|
||||
from langchain_openai import ChatOpenAI
|
||||
|
||||
from langchain_core.outputs import ChatGeneration, ChatGenerationChunk, ChatResult
|
||||
from langchain_openai import ChatOpenAI
|
||||
|
||||
from deerflow.models.message_ordering import normalize_chat_payload_messages
|
||||
|
||||
|
||||
class PatchedChatOpenAI(ChatOpenAI):
|
||||
"""ChatOpenAI with ``thought_signature`` preservation for Gemini thinking via OpenAI gateway.
|
||||
@@ -87,15 +89,16 @@ class PatchedChatOpenAI(ChatOpenAI):
|
||||
if payload_msg.get("role") == "assistant" and isinstance(orig_msg, AIMessage):
|
||||
_restore_tool_call_signatures(payload_msg, orig_msg)
|
||||
_restore_reasoning_content(payload_msg, orig_msg)
|
||||
else:
|
||||
# Fallback: match assistant-role entries positionally against AIMessages.
|
||||
ai_messages = [m for m in original_messages if isinstance(m, AIMessage)]
|
||||
assistant_payloads = [(i, m) for i, m in enumerate(payload_messages) if m.get("role") == "assistant"]
|
||||
for (_, payload_msg), ai_msg in zip(assistant_payloads, ai_messages):
|
||||
_restore_tool_call_signatures(payload_msg, ai_msg)
|
||||
_restore_reasoning_content(payload_msg, ai_msg)
|
||||
|
||||
return payload
|
||||
else:
|
||||
# Fallback: match assistant-role entries positionally against AIMessages.
|
||||
ai_messages = [m for m in original_messages if isinstance(m, AIMessage)]
|
||||
assistant_payloads = [(i, m) for i, m in enumerate(payload_messages) if m.get("role") == "assistant"]
|
||||
for (_, payload_msg), ai_msg in zip(assistant_payloads, ai_messages):
|
||||
_restore_tool_call_signatures(payload_msg, ai_msg)
|
||||
_restore_reasoning_content(payload_msg, ai_msg)
|
||||
|
||||
normalize_chat_payload_messages(payload)
|
||||
return payload
|
||||
|
||||
def _convert_chunk_to_generation_chunk(
|
||||
self,
|
||||
|
||||
+13
-10
@@ -32,9 +32,11 @@ from langchain_core.messages import (
|
||||
)
|
||||
from langchain_core.messages.tool import tool_call_chunk
|
||||
from langchain_core.outputs import ChatGeneration, ChatGenerationChunk, ChatResult
|
||||
from langchain_openai import ChatOpenAI
|
||||
from langchain_openai.chat_models.base import _create_usage_metadata
|
||||
|
||||
from langchain_openai import ChatOpenAI
|
||||
from langchain_openai.chat_models.base import _create_usage_metadata
|
||||
|
||||
from deerflow.models.message_ordering import normalize_chat_payload_messages
|
||||
|
||||
|
||||
def _normalize_vllm_chat_template_kwargs(payload: dict[str, Any]) -> None:
|
||||
"""Map DeerFlow's legacy ``thinking`` toggle to vLLM/Qwen's ``enable_thinking``.
|
||||
@@ -182,13 +184,14 @@ class VllmChatModel(ChatOpenAI):
|
||||
for payload_msg, orig_msg in zip(payload_messages, original_messages):
|
||||
if payload_msg.get("role") == "assistant" and isinstance(orig_msg, AIMessage):
|
||||
_restore_reasoning_field(payload_msg, orig_msg)
|
||||
else:
|
||||
ai_messages = [message for message in original_messages if isinstance(message, AIMessage)]
|
||||
assistant_payloads = [message for message in payload_messages if message.get("role") == "assistant"]
|
||||
for payload_msg, ai_msg in zip(assistant_payloads, ai_messages):
|
||||
_restore_reasoning_field(payload_msg, ai_msg)
|
||||
|
||||
return payload
|
||||
else:
|
||||
ai_messages = [message for message in original_messages if isinstance(message, AIMessage)]
|
||||
assistant_payloads = [message for message in payload_messages if message.get("role") == "assistant"]
|
||||
for payload_msg, ai_msg in zip(assistant_payloads, ai_messages):
|
||||
_restore_reasoning_field(payload_msg, ai_msg)
|
||||
|
||||
normalize_chat_payload_messages(payload)
|
||||
return payload
|
||||
|
||||
def _create_chat_result(self, response: dict | openai.BaseModel, generation_info: dict | None = None) -> ChatResult:
|
||||
"""Preserve vLLM reasoning on non-streaming responses."""
|
||||
|
||||
Reference in new issue
Block a user