.Net: Added IChatHistoryReducer for ChatCompletion with AzureOpenAI and OpenAI

[sandrohanea]

Motivation and Context

When using Auto-Invocation for Tools with OpenAI and AzureOpenAI, there is no way to intercept and reduce the chat history after tools are being invoked.

This is needed as some tools might return long responses and we might want to execute different prompt compression / optimization based on the tool result.

Note: AutoFunctionInvocationFilters cannot be used as they are invoked only before the result of the tools is added to the chat history.

The alternative to this approach would be to intercept the http call to Open AI and alter the request directly but that will cause multiple allocation, serializations and deserializations and won't scale to other Connectors.

Description

This change is extracting IChatHistoryReducer from Agents.Core project and moving it to SK.Abstrations.AI.ChatCompletions.

IChatHistoryReducer was renamed to IAgentChatHistoryReducer in order to avoid confusion as it requires 2 more methods: Equals and GetHashCode that are used to compute the agent channel key. It now inherits IChatHistoryReducer.

Added tests to ensure that a DI registered reducer is used in the ChatCompletionService of both OpenAI and AzureOpenAI connectors.

Important notice: the reducer is not mutating the ChatHistory, as it might be that new tool invocation or chat turns will require different reducing logic. (e.g. context-aware summarization)

Contribution Checklist

• The code builds clean without any errors or warnings
• The PR follows the SK Contribution Guidelines and the pre-submission formatting script raises no violations
• All unit tests pass, and I have added new tests where possible
• I didn't break anyone 😄

[crickman]

Would it make sense to provide both implementations of the reducer in the core framework:

• Truncation
• Summarization

[markwallace-microsoft]

Hi @sandrohanea thanks for creating this PR. The team met to discuss and a few things came up

1. We currently mutate ChatHistory when automatic function calling is enabled. A function calling loop can be long and have many round trips to the model and truncating the ChatHistory could have undesired side effects. In general the consensus was that we don't want to chat history reducer to be called by the chat completion service.
2. The code shows how to use the reducer with Dependency Injection but we also need to handle non DI cases.

The consensus was that developers have control over the ChatHistory that they send to the chat completion service so they can reduce it as necessary and also decide which messages from the response get added back into the ChatHistory. We're going to work on some samples (during the current sprint) for this we will share with you once we have them.

[sandrohanea]

Closing this PR as it's not currently aligning with SK plan for abstractions.

[qihangnet]

@markwallace-microsoft Can the Reducer implementation in #7570 be moved to Core? I believe it is useful not just for the Agent.