HNSW Graph: Design Specification

[Iamdavidonuh]

This PR begins the work on #184

[deven96]

We didn't seem to cover deletion of a node

[Iamdavidonuh]

Regarding deletion: the original HNSW paper does not describe any true node-deletion procedure, which leads me to assume full deletion isn’t supported. Given how interconnected the graph becomes, removing a node would effectively require rebuilding significant parts of the structure. The only related mechanism discussed is neighbor pruning through search-layer heuristics, not actual removal. We can mark full deletion as a TODO / requires research.

[deven96]

Not having deletion isnt particularly tenable though but I agree we can mark as TODO

In order to merge it in though we'd need to give users the assurance that upon removing embeddings and searching through the index doesn't turn up stale vectors

[Iamdavidonuh]

There's now an algorithm for delete, check it out.

[nnethercott]

if you're curious this paper from microsoft outlines a pretty solid deletion strategy which ensures your search won't degrade after multiple rounds of insertions/deletions

[Iamdavidonuh]

Thank you for this reference! I’ll check it out.

[Iamdavidonuh]

@nnethercott Thanks again for the reference! I went through it and it’s a solid approach for deletion.

For my implementation, I’m leaning toward a slightly different strategy: maintaining back-links for each node. This makes deletion straightforward, as we only need to remove references from incoming neighbors (back-links), keeping the operation localized. There’s a memory overhead, but for the speed and simplicity of deletions, it’s a trade-off I’m willing to accept for now.

I understand that this approach means insertions will involve more updates to other nodes, and I’m curious to see how costly that will be in practice. For now, I plan to stick with this approach and will consider reverting only if the drawbacks outweigh the benefits.

We’ll also test recall by potentially replicating the deletion experiment from the paper: randomly remove a percentage of nodes and reinsert them over multiple cycles, then observe whether search performance (recall) remains stable. This will allow us to compare how our back-link deletion strategy performs relative to the Delete Policies described in the article.