github microsoft/autogen v0.4.3
on GitHub

20 hours ago

What's new

This is the first release since 0.4.0 with significant new features! We look forward to hearing feedback and suggestions from the community.

Chat completion model cache

One of the big missing features from 0.2 was the ability to seamlessly cache model client completions. This release adds ChatCompletionCache which can wrap any other ChatCompletionClient and cache completions.

There is a CacheStore interface to allow for easy implementation of new caching backends. The currently available implementations are: 

import asyncio
import tempfile

from autogen_core.models import UserMessage
from autogen_ext.models.openai import OpenAIChatCompletionClient
from autogen_ext.models.cache import ChatCompletionCache, CHAT_CACHE_VALUE_TYPE
from autogen_ext.cache_store.diskcache import DiskCacheStore
from diskcache import Cache


async def main():
    with tempfile.TemporaryDirectory() as tmpdirname:
        openai_model_client = OpenAIChatCompletionClient(model="gpt-4o")

        cache_store = DiskCacheStore[CHAT_CACHE_VALUE_TYPE](Cache(tmpdirname))
        cache_client = ChatCompletionCache(openai_model_client, cache_store)

        response = await cache_client.create([UserMessage(content="Hello, how are you?", source="user")])
        print(response)  # Should print response from OpenAI
        response = await cache_client.create([UserMessage(content="Hello, how are you?", source="user")])
        print(response)  # Should print cached response


asyncio.run(main())

ChatCompletionCache is not yet supported by the declarative component config, see the issue to track progress.

#4924 by @srjoglekar246

GraphRAG

This releases adds support for GraphRAG as a tool agents can call. You can find a sample for how to use this integration here, and docs for LocalSearchTool and GlobalSearchTool. 

import asyncio
from autogen_ext.models.openai import OpenAIChatCompletionClient
from autogen_agentchat.ui import Console
from autogen_ext.tools.graphrag import GlobalSearchTool
from autogen_agentchat.agents import AssistantAgent


async def main():
    # Initialize the OpenAI client
    openai_client = OpenAIChatCompletionClient(
        model="gpt-4o-mini",
    )

    # Set up global search tool
    global_tool = GlobalSearchTool.from_settings(settings_path="./settings.yaml")

    # Create assistant agent with the global search tool
    assistant_agent = AssistantAgent(
        name="search_assistant",
        tools=[global_tool],
        model_client=openai_client,
        system_message=(
            "You are a tool selector AI assistant using the GraphRAG framework. "
            "Your primary task is to determine the appropriate search tool to call based on the user's query. "
            "For broader, abstract questions requiring a comprehensive understanding of the dataset, call the 'global_search' function."
        ),
    )

    # Run a sample query
    query = "What is the overall sentiment of the community reports?"
    await Console(assistant_agent.run_stream(task=query))


if __name__ == "__main__":
    asyncio.run(main())

#4612 by @lspinheiro

Semantic Kernel model adapter

Semantic Kernel has an extensive collection of AI connectors. In this release we added support to adapt a Semantic Kernel AI Connector to an AutoGen ChatCompletionClient using the SKChatCompletionAdapter.

Currently this requires passing the kernel during create, and so cannot be used with AssistantAgent directly yet. This will be fixed in a future release (#5144).

#4851 by @lspinheiro

AutoGen to Semantic Kernel tool adapter

We also added a tool adapter, but this time to allow AutoGen tools to be added to a Kernel, called KernelFunctionFromTool.

#4851 by @lspinheiro

Jupyter Code Executor

This release also brings forward Jupyter code executor functionality that we had in 0.2, as the JupyterCodeExecutor.

Please note that this currently on supports local execution and should be used with caution.

#4885 by @Leon0402

Memory

It's still early on but we merged the interface for agent memory in this release. This allows agents to enrich their context from a memory store and save information to it. The interface is defined in core and AssistantAgent in agentchat accepts memory as a parameter now. There is an initial example memory implementation which simply injects all memories as system messages for the agent. The intention is for the memory interface to be able to be used for both RAG and agent memory systems going forward.

#4438 by @victordibia, #5053 by @ekzhu

Declarative config

We're continuing to expand support for declarative configs throughout the framework. In this release, we've added support for termination conditions and base chat agents. Once we're done with this, you'll be able to configure and entire team of agents with a single config file and have it work seamlessly with AutoGen studio. Stay tuned!

#4984, #5055 by @victordibia

Other

  • Add sources field to TextMentionTermination by @Leon0402 in #5106
  • Update gpt-4o model version to 2024-08-06 by @ekzhu in #5117

Bug fixes

  • Retry multiple times when M1 selects an invalid agent. Make agent sel… by @afourney in #5079
  • fix: normalize finish reason in CreateResult response by @ekzhu in #5085
  • Pass context between AssistantAgent for handoffs by @ekzhu in #5084
  • fix: ensure proper handling of structured output in OpenAI client and improve test coverage for structured output by @ekzhu in #5116
  • fix: use tool_calls field to detect tool calls in OpenAI client; add integration tests for OpenAI and Gemini by @ekzhu in #5122

Other changes

New Contributors

Full Changelog: v0.4.1...v0.4.3

Check out latest releases or
releases around microsoft/autogen v0.4.3

Don't miss a new autogen release

NewReleases is sending notifications on new releases.

Get notifications