●Change8

v0.4.3

📦 autogen
✨ 10 features🐛 4 fixes🔧 14 symbols

Summary

This release introduces significant features including model client caching, GraphRAG integration, Semantic Kernel adapters, and a new agent memory interface. It also expands declarative configuration support and brings back the Jupyter code executor.

Migration Steps

  1. To use caching, wrap your ChatCompletionClient with ChatCompletionCache and provide a CacheStore implementation.
  2. If using JupyterCodeExecutor, ensure it is used in a secure local environment as it currently only supports local execution.
  3. Update AssistantAgent implementations to utilize the new memory parameter if context enrichment is required.

✨ New Features

  • Added ChatCompletionCache to wrap model clients and cache completions.
  • Introduced CacheStore interface with DiskCacheStore and RedisStore implementations.
  • Added GraphRAG support via LocalSearchTool and GlobalSearchTool.
  • Added SKChatCompletionAdapter to support Semantic Kernel AI connectors as AutoGen model clients.
  • Added KernelFunctionFromTool to adapt AutoGen tools for use within Semantic Kernel.
  • Introduced JupyterCodeExecutor for local code execution.
  • Added agent memory interface and support for memory parameters in AssistantAgent.
  • Expanded declarative configuration support for termination conditions and base chat agents.
  • Added sources field to TextMentionTermination.
  • Updated default gpt-4o model version to 2024-08-06.

🐛 Bug Fixes

  • Improved reliability by retrying multiple times when M1 selects an invalid agent.
  • Normalized finish reason in CreateResult responses.
  • Fixed context passing between AssistantAgent instances during handoffs.
  • Ensured proper handling of structured output in the OpenAI client.

🔧 Affected Symbols

ChatCompletionCacheChatCompletionClientCacheStoreDiskCacheStoreRedisStoreLocalSearchToolGlobalSearchToolSKChatCompletionAdapterKernelFunctionFromToolJupyterCodeExecutorMemoryAssistantAgentTextMentionTerminationCreateResult