feat: enhance request payload handling for DeepSeek API integration

Author: liuruibin

apps/models_provider/impl/base_chat_open_ai.py

@@ -1,30 +1,31 @@
 # coding=utf-8
 import base64
-import json
 from concurrent.futures import ThreadPoolExecutor
-from requests.exceptions import ConnectTimeout, ReadTimeout
 from typing import Dict, Optional, Any, Iterator, cast, Union, Sequence, Callable, Mapping
 
 from langchain_core.language_models import LanguageModelInput
 from langchain_core.messages import BaseMessage, get_buffer_string, BaseMessageChunk, HumanMessageChunk, AIMessageChunk, \
     SystemMessageChunk, FunctionMessageChunk, ChatMessageChunk
-from langchain_core.messages.ai import UsageMetadata, AIMessage
+from langchain_core.messages.ai import UsageMetadata
 from langchain_core.messages.tool import tool_call_chunk, ToolMessageChunk
 from langchain_core.outputs import ChatGenerationChunk
 from langchain_core.runnables import RunnableConfig, ensure_config
 from langchain_core.tools import BaseTool
 from langchain_openai import ChatOpenAI
 from langchain_openai.chat_models.base import _create_usage_metadata
+from requests.exceptions import ReadTimeout
 
 from common.config.tokenizer_manage_config import TokenizerManage
 from common.utils.logger import maxkb_logger
 
+
 def custom_get_token_ids(text: str):
     tokenizer = TokenizerManage.get_tokenizer()
     return tokenizer.encode(text)
 
+
 def _convert_delta_to_message_chunk(
-    _dict: Mapping[str, Any], default_class: type[BaseMessageChunk]
+        _dict: Mapping[str, Any], default_class: type[BaseMessageChunk]
 ) -> BaseMessageChunk:
     """Convert to a LangChain message chunk."""
     id_ = _dict.get("id")
@@ -80,6 +81,7 @@ def _convert_delta_to_message_chunk(
         return ChatMessageChunk(content=content, role=role, id=id_)
     return default_class(content=content, id=id_)  # type: ignore[call-arg]#
 
+
 class BaseChatOpenAI(ChatOpenAI):
     usage_metadata: dict = {}
     custom_get_token_ids = custom_get_token_ids
@@ -219,64 +221,13 @@ def invoke(
             'token_usage'] if 'token_usage' in chat_result.response_metadata else chat_result.usage_metadata
         return chat_result
 
-    def _get_request_payload(
-            self,
-            input_: LanguageModelInput,
-            *,
-            stop: list[str] | None = None,
-            **kwargs: Any,
-    ) -> dict:
-        # Get original messages to preserve reasoning_content before base conversion
-        messages = self._convert_input(input_).to_messages()
-        # Store reasoning_content for AIMessages with tool_calls
-        # According to DeepSeek API docs, reasoning_content is REQUIRED when tool_calls
-        # are present during the tool invocation process (within same question/turn).
-        # See: https://api-docs.deepseek.com/guides/thinking_mode#tool-calls
-        reasoning_content_map = {}
-        for i, msg in enumerate(messages):
-            if (
-                    isinstance(msg, AIMessage)
-                    and (msg.tool_calls or msg.invalid_tool_calls)
-                    and (reasoning := msg.additional_kwargs.get("reasoning_content"))
-            ):
-                reasoning_content_map[i] = reasoning
-
-        payload = super()._get_request_payload(input_, stop=stop, **kwargs)
-
-        # Restore reasoning_content for assistant messages with tool_calls
-        # This is required by DeepSeek API - missing it causes 400 error
-        if "messages" in payload and reasoning_content_map:
-            for i, message in enumerate(payload["messages"]):
-                if (
-                        i in reasoning_content_map
-                        and message.get("role") == "assistant"
-                        and message.get("tool_calls")
-                ):
-                    message["reasoning_content"] = reasoning_content_map[i]
-
-        # Apply DeepSeek-specific message formatting
-        for message in payload["messages"]:
-            if message["role"] == "tool" and isinstance(message["content"], list):
-                message["content"] = json.dumps(message["content"])
-            elif message["role"] == "assistant" and isinstance(
-                    message["content"], list
-            ):
-                # DeepSeek API expects assistant content to be a string, not a list.
-                # Extract text blocks and join them, or use empty string if none exist.
-                text_parts = [
-                    block.get("text", "")
-                    for block in message["content"]
-                    if isinstance(block, dict) and block.get("type") == "text"
-                ]
-                message["content"] = "".join(text_parts) if text_parts else ""
-        return payload
-
     def upload_file_and_get_url(self, file_stream, file_name):
         """上传文件并获取文件URL"""
         base64_video = base64.b64encode(file_stream).decode("utf-8")
         video_format = get_video_format(file_name)
         return f'data:{video_format};base64,{base64_video}'
 
+
 def get_video_format(file_name):
     extension = file_name.split('.')[-1].lower()
     format_map = {
@@ -285,4 +236,4 @@ def get_video_format(file_name):
         'mov': 'video/mov',
         'wmv': 'video/x-ms-wmv'
     }
-    return format_map.get(extension, 'video/mp4')
+    return format_map.get(extension, 'video/mp4')

apps/models_provider/impl/deepseek_model_provider/model/llm.py

@@ -6,7 +6,11 @@
 @Author  ：Brian Yang
 @Date    ：5/12/24 7:44 AM 
 """
-from typing import Dict
+import json
+from typing import Dict, Any
+
+from langchain_core.language_models import LanguageModelInput
+from langchain_core.messages import AIMessage
 
 from models_provider.base_model_provider import MaxKBBaseModel
 from models_provider.impl.base_chat_open_ai import BaseChatOpenAI
@@ -29,3 +33,55 @@ def new_instance(model_type, model_name, model_credential: Dict[str, object], **
             extra_body=optional_params
         )
         return deepseek_chat_open_ai
+
+    def _get_request_payload(
+            self,
+            input_: LanguageModelInput,
+            *,
+            stop: list[str] | None = None,
+            **kwargs: Any,
+    ) -> dict:
+        # Get original messages to preserve reasoning_content before base conversion
+        messages = self._convert_input(input_).to_messages()
+        # Store reasoning_content for AIMessages with tool_calls
+        # According to DeepSeek API docs, reasoning_content is REQUIRED when tool_calls
+        # are present during the tool invocation process (within same question/turn).
+        # See: https://api-docs.deepseek.com/guides/thinking_mode#tool-calls
+        reasoning_content_map = {}
+        for i, msg in enumerate(messages):
+            if (
+                    isinstance(msg, AIMessage)
+                    and (msg.tool_calls or msg.invalid_tool_calls)
+                    and (reasoning := msg.additional_kwargs.get("reasoning_content"))
+            ):
+                reasoning_content_map[i] = reasoning
+
+        payload = super()._get_request_payload(input_, stop=stop, **kwargs)
+
+        # Restore reasoning_content for assistant messages with tool_calls
+        # This is required by DeepSeek API - missing it causes 400 error
+        if "messages" in payload and reasoning_content_map:
+            for i, message in enumerate(payload["messages"]):
+                if (
+                        i in reasoning_content_map
+                        and message.get("role") == "assistant"
+                        and message.get("tool_calls")
+                ):
+                    message["reasoning_content"] = reasoning_content_map[i]
+
+        # Apply DeepSeek-specific message formatting
+        for message in payload["messages"]:
+            if message["role"] == "tool" and isinstance(message["content"], list):
+                message["content"] = json.dumps(message["content"])
+            elif message["role"] == "assistant" and isinstance(
+                    message["content"], list
+            ):
+                # DeepSeek API expects assistant content to be a string, not a list.
+                # Extract text blocks and join them, or use empty string if none exist.
+                text_parts = [
+                    block.get("text", "")
+                    for block in message["content"]
+                    if isinstance(block, dict) and block.get("type") == "text"
+                ]
+                message["content"] = "".join(text_parts) if text_parts else ""
+        return payload