.Net: Text Search Service ADR #6012
Conversation
markwallace-microsoft
added
.NET
github-actions
bot
changed the title
github-actions
bot
changed the title
markwallace-microsoft
changed the title
markwallace-microsoft
removed
the
python
markwallace-microsoft
force-pushed
the
users/markwallace/text-search-service-adr
branch
from
13b1f6f
to
2bae9e8
Compare
…nto users/markwallace/text-search-service-adr
…nto users/markwallace/text-search-service-adr
markwallace-microsoft
changed the title
| ## Context and Problem Statement | ||
|
|
||
| Semantic Kernel has support for searching using popular Vector databases e.g. Azure AI Search, Chroma, Milvus and also Web search engines e.g. Bing, Google. | ||
| There are two sets of abstractions and plugins depending on whether the developer wants to perform search against a Vector database of a Web search engine. |
There was a problem hiding this comment.
Should it be "a Vector database or a Web search engine."?
| var stringPlugin = new TextSearchPlugin<string>(searchService); | ||
| kernel.ImportPluginFromObject(stringPlugin, "StringSearch"); | ||
| var pagePlugin = new TextSearchPlugin<BingWebPage>(searchService); | ||
| kernel.ImportPluginFromObject(pagePlugin, "PageSearch"); | ||
|
|
||
| // Invoke the plugin to perform a text search and return string values | ||
| var question = "What is the Semantic Kernel?"; | ||
| var function = kernel.Plugins["StringSearch"]["Search"]; | ||
| var result = await kernel.InvokeAsync(function, new() { ["query"] = question }); | ||
|
|
||
| Console.WriteLine(result); | ||
|
|
||
| // Invoke the plugin to perform a text search and return BingWebPage values | ||
| function = kernel.Plugins["PageSearch"]["Search"]; | ||
| result = await kernel.InvokeAsync(function, new() { ["query"] = question, ["count"] = 2 }); |
There was a problem hiding this comment.
SearchAsync method is generic, and each implementation has these checks: typeof(T) == typeof(string) or typeof(T) == typeof(TextSearchResult). TextSearchPlugin class is also generic, and for each specific type you need to create separate instance (e.g. TextSearchPlugin<string> and TextSearchPlugin<BingWebPage>).
My point during initial discussion was - with generics and type checking it's already a sacrifice, but at the end of the day we still can get only string from kernel.InvokeAsync, and if it's object, it will be serialized. I think it will be really useful to get the actual search result type like result.GetValue<BingWebPage> after InvokeAsync method. Otherwise, we can probably avoid using generics, work just with object, and the result of this sample will still be the same.
Just some ideas what we can do in theory:
- Make
ITextSearchService<T>generic. - For specific service (e.g.
BingTextSearchService), implement interface with all possible types that can be produced by this service:BingTextSearchService: ITextSearchService<string>, ITextSearchService<BingWebPage>- this should allow to avoidtypeof(T) == typeof(string)check inSearchAsyncimplementation. - Based on that, we don't need
TextSearchPluginclass, because we can mark allSearchAsyncmethods inBingTextSearchServiceasKernelFunctionand name them as you did in line 27 and 29 -StringSearchandPageSearch. We will end up withBingPlugin-StringSearchandBingPlugin-PageSearchin kernel. If we share it with LLM, it will be able to decide which function to use based on user prompt. After specific function is invoked, it will return specific type, which will be possible to get withresult.GetValue<T>.
Another way for point 3, we can keep TextSearchPlugin as non-generic, which will accept ITextSearchService<string> and/or ITextSearchService<WebPage> in constructor. Then, it will be possible to inject any service that implements ITextSearchService<string> and/or ITextSearchService<WebPage> (e.g. Bing or Google). In this case, you won't need to initialize TextSearchPlugin multiple times, DI can resolve all dependencies for you, there will be no need in type checking in search methods and complex type should be available as result of InvokeAsync method.
Let me know what you think about it.
Motivation and Context
This is a POC for the new text search service abstraction. Sharing very early for team collaboration.
Description
Contribution Checklist