.Net: Example of Semantic Caching with Filters

[dmytrostruk]

Motivation and Context

This example shows how to achieve Semantic Caching with Filters.
IPromptRenderFilter is used to get rendered prompt and check in cache if similar prompt was already answered. If there is a record in cache, then previously cached answer will be returned to the user instead of making a call to LLM. If there is no record in cache, a call to LLM will be performed, and result will be cached together with rendered prompt.
IFunctionInvocationFilter is used to update cache with rendered prompt and related LLM result.

Example includes in-memory, Redis and Azure Cosmos DB for MongoDB as caching stores.
Common output which demonstrates that second execution is faster, because the result is returned from cache:

First run: What's the tallest building in New York?
Elapsed Time: 00:00:03.828
Second run: What is the highest building in New York City?
Elapsed Time: 00:00:00.541
Result 1: The tallest building in New York is One World Trade Center, also known as Freedom Tower. It stands at 1,776 feet (541.3 meters) tall, including its spire.
Result 2: The tallest building in New York is One World Trade Center, also known as Freedom Tower. It stands at 1,776 feet (541.3 meters) tall, including its spire.

PR also contains a couple of fixes in Azure Cosmos DB for MongoDB connector and a couple of additions in public API:

1. Added FunctionResult? Result property to PromptRenderContext. By default it's null, because at prompt rendering stage there is no available result yet. But it's possible to set result with some value - in this case, prompt won't be sent to LLM. Instead, the result from filter will be returned.
2. Added string? RenderedPrompt to FunctionResult type as Experimental. By default it's null, and will be populated only when KernelFunctionFromPrompt is executed. This property will provide a couple of benefits:
   • It's an additional way how to observe rendered prompt which was sent to LLM during function invocation (today, it's possible to see it only through filter or trace logging).
   • Rendered prompt will be also available in function invocation/automatic function invocation filters, which is required for caching scenarios to store rendered prompt and LLM result together.

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 😄