Access to Chat history in semantic kernel function #7186

That is correct Dmytro Struk (@dmytrostruk). This is our UC:

We have a semantic function that searches Azure AI search and returns a set of relevant articles back to the LLM - basically a classic RAG scenario. This is done as an SK function as this is just one of the things a user might ask, and potentially, a user can say:

Give 10 latest articles about XYZ and put them into a PDF document:

In this case, we would invoke our Azure AI search, followed by an SK function that generates PDFs.

The issue here is that the SK function for Azure AI is quite complex; we build a filter with 10ish parameters (some of the parameters can be data range, content type, authors, keyword, order, number of results ...)

So what we have on the SK function is:

1. Parsing user's request to extract all search parameters -> for this we use a call to LLM
2. Creating Azure AI search filter
3. Executing Azure AI search
4. Postprocessing Azure AI search results
5. return a list of results back

Our main semantic function accepts user requests as a parameter, which we pass to the SK function's LLM to extract search parameters (no 1).

When we have a conversation with multiple back and forth e.g.:

Call 1:
Give 10 latest articles about XYZ and put them into a PDF document:
Call 2 can you focus on just articles published in last 2 months
Call 3 filter also by author XXX

The input we get during call 3 usually doesn't contain everything the user is looking for. So this is the moment where we need full history access inside SK function so we can pass on full history to the LLM to be able to extra all the parameters we use for Azure Ai Search:
no of result: 10
content-type: articles
keyword: XYZ
date range today- 2 months
Authror= XXX

hope this explains

thanks

JP-droidic Got it, thanks for additional information! Have you considered to maintain ChatHistory on your side and invoke your requests in a loop? Like in this example:

This is just a demonstration, if you have a web server, you can serialize and store chat history, and then re-use it for further requests, but the main point here, based on your example, you should be able to achieve it if you store chat history on your side and then re-use it with additional requests, so the model will better understand the context. It's still possible to enable auto function calling and invoke Azure AI Search related operations, that are wrapped in SK functions, if needed. Please let us know if it's something that can handle your scenario, thanks!

yes, that makes sense. Thank you for your help