Single-Index Metadata Filtering

Hi all,

I have been using the Pinecone metadata filters by using the Langchain Self-Query Retriever. This has been really helpful for applying hard filters on my inventory while also doing embedding similarity in one query.

One issue I have run into is applying the hard filters on string attributes. My datasets tagging can be a bit variable. For example, if I have a piece of clothing inventory with an attribute style, the values can vary but have the same meaning. One example is an item could be tagged as Style == button-down or Style == button downs. I realize that there are ways to improve my attribute tagging; however, I am wondering if there is any chance that a filtering capability for strings that does prefix/suffix matching and/or does a similarity match is in the works (if possible).

Any input on this issue including alternate ways to address this problem is much appreciated.


After doing some more research, I am seeing some methods such as word2vec similarity that could just be applied on the filter input and the metadata field values. Additional methods include fuzzy matching and proximity features. Any plans to support these types of filters?