klbr/khoj
1
0
Fork 0
mirror of https://github.com/khoaliber/khoj.git synced 2026-03-07 21:29:13 +00:00
Code Issues Packages Projects Releases Wiki Activity

Parse DeepSeek reasoning model thoughts served via OpenAI compatible API

Browse Source 
DeepSeek reasoners returns reasoning in reasoning_content field.

Create an async stream processor to parse the reasoning out when using
the deepseek reasoner model.
This commit is contained in:
Debanjum
2025-05-02 18:36:24 -06:00
parent 16f3c85dde
commit 61a50efcc3
1 changed files with 14 additions and 1 deletions
Download Patch File Download Diff File Expand all files Collapse all files

15
src/khoj/processor/conversation/openai/utils.py
View File

@@ -184,6 +184,7 @@ async def chat_completion_with_backoff(
reasoning_effort = "high" if deepthought else "low" reasoning_effort = "high" if deepthought else "low"
model_kwargs["reasoning_effort"] = reasoning_effort model_kwargs["reasoning_effort"] = reasoning_effort
elif model_name.startswith("deepseek-reasoner"): elif model_name.startswith("deepseek-reasoner"):
stream_processor = adeepseek_stream_processor
# Two successive messages cannot be from the same role. Should merge any back-to-back messages from the same role. # Two successive messages cannot be from the same role. Should merge any back-to-back messages from the same role.
# The first message should always be a user message (except system message). # The first message should always be a user message (except system message).
updated_messages: List[dict] = [] updated_messages: List[dict] = []
@@ -194,7 +195,6 @@ async def chat_completion_with_backoff(
updated_messages[-1]["content"] += " " + message["content"] updated_messages[-1]["content"] += " " + message["content"]
else: else:
updated_messages.append(message) updated_messages.append(message)
formatted_messages = updated_messages formatted_messages = updated_messages
elif is_qwen_reasoning_model(model_name, api_base_url): elif is_qwen_reasoning_model(model_name, api_base_url):
stream_processor = partial(ain_stream_thought_processor, thought_tag="think") stream_processor = partial(ain_stream_thought_processor, thought_tag="think")
@@ -349,6 +349,19 @@ async def adefault_stream_processor(
yield ChatCompletionWithThoughtsChunk.model_validate(chunk.model_dump()) yield ChatCompletionWithThoughtsChunk.model_validate(chunk.model_dump())
async def adeepseek_stream_processor(
chat_stream: openai.AsyncStream[ChatCompletionChunk],
) -> AsyncGenerator[ChatCompletionWithThoughtsChunk, None]:
"""
Async generator to cast and return chunks from the deepseek chat completions stream.
"""
async for chunk in chat_stream:
tchunk = ChatCompletionWithThoughtsChunk.model_validate(chunk.model_dump())
if len(tchunk.choices) > 0 and tchunk.choices[0].delta.reasoning_content:
tchunk.choices[0].delta.thought = chunk.choices[0].delta.reasoning_content
yield tchunk
def in_stream_thought_processor( def in_stream_thought_processor(
chat_stream: openai.Stream[ChatCompletionChunk], thought_tag="think" chat_stream: openai.Stream[ChatCompletionChunk], thought_tag="think"
) -> Generator[ChatCompletionStreamWithThoughtEvent, None, None]: ) -> Generator[ChatCompletionStreamWithThoughtEvent, None, None]: