Below is what llama index provides for

i'm not 100% sure how the langchain retriever works

Does it return text chunks that match a given query? Then a custom retriever makes sense

We could maybe add wrappers for langchain stuff, but I'm curious why you arent using llama-index for retrieval ?

filters

Here's langchain's output from a OpenSearchVectorSearch object

It looks like llama-index almost supports at least one kind of filter though. Here's llama-index's query

Here is langchain's

I'm just going to stuff something in there and remove the validation about filtering being unimplemented and see what happens

Definitely worth noting though that llama-index is behind langchain in OpenSearch support.

Another reason to use it is that AWS has an OpenSearch service, and afaik it's the only vector store db I can use while keeping my company's legal and security depts happy. (edit: and satisfies my other requirements)

@Logan M I did get a basic boolean filter to work by small edits to llama-index/vector_stores/opensearch.py, but comparing functionality in more depth I think a wrapper is a better option until llama-index implements something more sophisticated.

Yea the vector store integrations are mostly community driven. Feel free to make a PR. Sadly opensearch is barely used (at least judging from discord/github issues), so it's a little barebones at the moment

As it turns out, OpenSearch's KNN filtering is applied after the k-results are retrieved anyway, and as such it would be just about as easy to filter the response instead of asking for a filtered response. Their Lucene engine has pre-filtering, but it only supports up to dimension 1024

Or alternatively, there is a brute force exact Knn that lets you pre-filter, but doesn't scale well

More details https://opensearch.org/docs/latest/search-plugins/knn/filter-search-knn/

tl;dr exact match filtering might not behave as expected unless you are using "Script Scoring" or "Painless Scripting", but those are not as scalable / flexible as the approximate-Knn to which only a "boolean" filter may be applied.

I submitted a PR to support filtering, see #🙌contributing