fix(adapters): handle non-stream content with reasoning for responses#2885
fix(adapters): handle non-stream content with reasoning for responses#2885xhebox wants to merge 1 commit intoolimorris:mainfrom
Conversation
format Signed-off-by: xhe <xw897002528@gmail.com>
|
Can you share what model you've encountered this with? The PR description is lacking detail. |
Deepseek behind an LLM gateway called axonhub. But I think it is general, the fake test case is exactly same as the deepseek api. However it did not test the content at all. EDIT: openai has the similar example too, https://developers.openai.com/api/docs/guides/migrate-to-responses. I met the problem when I use |
| h.expect_contains("**Choosing descriptive terms**", output.reasoning.content) | ||
| h.eq("rs_0a10a8c968d594670168e91d0204ac8195b26b3e4de997f65c", output.reasoning.id) | ||
| h.eq("gAAAAABo6", output.reasoning.encrypted_content) | ||
| h.eq("Dynamic expressive", output.content) |
There was a problem hiding this comment.
FYI, sample data is not fully tested. Before the change, output.content is nil, while it is "Dynamic expressive" actually. The first block can be item.type == reasoning.
There was a problem hiding this comment.
I have no idea what you mean by "not fully tested".
I haven't tried this PR out yet but will run it through OpenAI's models at some point this weekend.
There was a problem hiding this comment.
I have no idea what you mean by "not fully tested".
mock data tests/adapters/http/stubs/openai_responses_inline.txt") has two items, type = reasoning for the first, type = message for the second.
We tested the reasoning content but not the message content in this test. It would not pass if we do check the message content.
|
Sorry for not coming back to this sooner. Can you give me a set of steps to test this? |
open a chat buffer with reasoning-enabled model(like deepseek-reasoner), make some random chats, then Bug will be trigerred if the response has both reasoning and normal text content. |
|
Any news? I've tested it on a lot of other APIs like nvidia or GLM. |
2 participants
format
Description
Output text is not necessarily the first block. Deepseek-reasoner met this.
AI Usage
Related Issue(s)
Screenshots
Checklist
make allto ensure docs are generated, tests pass and StyLua has formatted the code