community: ClickHouse: Make it possible to not specify a vector index

[mneedham]

Vector indexes in ClickHouse are experimental at the moment and can sometimes break/change behaviour. So this PR makes it possible to say that you don't want to specify an index type.

Any queries against the embedding column will be brute force/linear scan, but that gives reasonable performance for small-medium dataset sizes.

In the other PR (#17247), @efriis asked:

> index_type: Optional[str] = "annoy"

What is this supposed to do? I think this will have unintended consequences in some string substitution stuff lower in the file

This makes index_type optional. I think I've defended against it being None inside the _schema function, which is where it's used. But let me know if you think I've missed somewhere else where it's used.

[vercel]

The latest updates on your projects. Learn more about Vercel for Git ↗︎

1 Ignored Deployment

Name	Status	Preview	Comments	Updated (UTC)
langchain	⬜️ Ignored (Inspect)	Visit Preview		Mar 13, 2024 9:51am

[efriis]

@mneedham just confirming - I'm pretty sure this change turns this into self.client.command(f"SET allow_experimental_None_index=1") unless I'm mistaken. This seems undesirable?

[mneedham]

@efriis I think I'm handling the empty index_type now, but it was also telling me that the width of the docstring that I added for the _schema function was too wide. I think I've fixed that, but not sure how to get it to run the linter on CI again?

[mneedham]

@efriis how do I get the build to run? I think I've fixed the lint issues that were there before.

[efriis]

Fixing in #19527

In the future if you make PRs from a personal branch, can push directly to your PR branch! https://python.langchain.com/docs/contributing/faq#how-do-i-allow-maintainers-to-edit-my-pr