| Age | Commit message (Collapse) | Author | Lines |
|
Propagate the per-request Idempotency-Key (generated once in the retry
loop) through request construction so every retry attempt reuses the same
key rather than minting a new one.
- Add field to ChatCompletionsData.
- Generate a single key per logical request in chat_completions and
chat_completions_streaming.
- Thread the key through request_builder (macro-based clients), Bedrock,
and VertexAI so the header is attached once and reused on retries.
|
|
- Add boolean to client config. When enabled, a
fresh (UUID) header is sent with each logical request
attempt so the outer retry loop reuses the same submission instead of
starting duplicate work on providers that support it (e.g. Sail Research).
The key is generated once per logical request (before the retry loop) and
reused across all retry attempts via the cloned ChatCompletionsData/request
builders, so providers debounce duplicate generations.
- Wire the header in request_builder (macro-based clients) and Bedrock's
chat_completions_builder.
- Document the option in config.example.yaml.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
- check the model type while retrieve model
- select chat/reranker model even if it is missed in client models
- find predefined-models for openai-compatible client with startsWith
- remove client::ApiType
|
|
|
|
- rename MessageContent:ToolResults to MessageContent:ToolCalls
- rename ToolResults to MessageContentToolCalls
- persist tool_calls to messages.md
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
* feat: support RAG
* support more embeddings models and implement concurrent embedding api
* show the progress of addings paths
* ignore embedding context when saving message
* embedding model max_chunk_size => default_chunk_size
* support pdf and pandoc formats (docx, epub, ipynb)
|
|
* rename `Completeion*` to `ChatCompletions*`
* rename `send_message*` to `chat_completions*`
* rename `request_builder` to `chat_completions_builder`
* rename `build_body` to `build_chat_completions_body`
* rename `extract_completion` to `extract_chat_completions`
* format
* remove unused config fields
|
|
|
|
* Removed extraneous key [stream] for AWS Bedrock Claude models.
* Reduceddefault AWS Bedrock llama-3 max_output_tokens to 2048 to align with API requirements.
---------
Co-authored-by: Rolf Wilms <rwilms@csc.com>
|
|
|
|
* feat: support function calling
* fix on Windows OS
* implement multi-steps function calling
* fix on Windows OS
* add error for client not support function calling
* refactor message data structure and make claude client supporting function calling
* support reuse previous call results
* improve error handling for function calling
* use prefix `may_` as indicator for `execute` type fucntions
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|