Add DeepSeek integration for Zed Assistant #23551
Conversation
- Add DeepSeek as a new language model provider - Implement DeepSeek API client with streaming support - Add DeepSeek configuration UI and settings - Update documentation with DeepSeek setup instructions - Add DeepSeek icon and model definitions - Integrate DeepSeek into language model registry
cla-bot
bot
added
the
cla-signed
maxdeviant
changed the title
This comment was marked as resolved.
This comment was marked as resolved.
This comment was marked as resolved.
This comment was marked as resolved.
- Add message merging to fix '400 Bad Request' error for Reasoner model
- Address error: 'deepseek-reasoner does not support successive user or assistant messages'
- Merge consecutive messages of same role before sending to API
- Handle User, System, Assistant and Tool message types
- Preserve tool_calls in merged assistant messages
- Add comprehensive test coverage for message merging logic
- Optimize string concatenation with reserve and push_str
- Make Model::id const for better performance
Fixes ERROR assistant_context_editor] Failed to connect to DeepSeek API: 400 Bad Request {'error':{'message':'deepseek-reasoner does not support successive user or assistant messages (messages[0] and messages[1] in your input). You should interleave the user/assistant messages in the message sequence.','type':'invalid_request_error','param':null,'code':'invalid_request_error'}}
|
Fixed issues: However, I'm not sure if the chosen method is the most appropriate. Welcome to discuss |
…ek API - Removed custom DeepSeekError struct and simplified error handling in the `complete` function. - Refactored `merge_consecutive_messages` to use a more straightforward iteration approach. - Introduced helper functions `merge_messages` and `can_merge` to improve readability. - Updated tests to reflect changes in message merging behavior.
- Added `Custom` variant to `deepseek::Model` enum to support custom models - Updated `id()`, `display_name()`, `max_token_count()`, and `max_output_tokens()` methods to handle custom models - Modified `provided_models()` in DeepSeek provider to include custom models from settings
|
I think the current implementation is ready for review. The only drawback is that the thinking process of |
… unused complete function - Removed the `complete` function and its associated tests from `deepseek.rs` - Moved message merging logic from `deepseek.rs` to `request.rs`'s `into_deepseek` method - Simplified `stream_completion` by removing redundant message merging - Updated message merging to handle both user and assistant messages in one pass - Maintained existing functionality while improving code organization
|
@Cupnfish you mention that this PR is "enabling users to utilize DeepSeek models for code completion". Which part of the changes deals with code completions in particular? I can't find anything implementing |
|
@shishkin Sorry for the misunderstanding, deepseek does indeed support code completion, but it was not implemented at the PR. I have already revised the release notes. If anyone is interested in implementing this feature, they can refer to this document: https://api-docs.deepseek.com/guides/fim_completion |
|
hi, |
Release Notes: