crclark/graph-anns

Efficient approximate nearest neighbor search data structure

graph-anns's Issues

license

Split up lib.rs

Find a more logical module structure

CI

Run

• tests
• lint
• formatting check
• clippy

parametrize node id type

To allow for graphs bigger than max value of u32

builder-style configs?

Right now configs are just structs. What's the best way to construct config objects in Rust such that adding new fields in the future doesn't break existing code?

Name it better

graph-anns is too generic. Maybe emphasize its genericity or incrementality.

serialization

Serialize and deserialize a graph and ensure it works

Try to adopt FINGER

https://www.amazon.science/blog/more-efficient-approximate-nearest-neighbor-search

Ensure we are using the right type for the distance callback

I think FnMut might be a better choice so the user can e.g. perform caching.

State machine testing

Use this or some other library so that we can test randomized sequences of insert/delete operations to ensure invariants hold.

Put prng into config instead of using thread_rng

compare to instant-distance

https://crates.io/crates/instant-distance

verify generated docs look sane

squash history and create new repo; don't use this one

Eliminate nightly requirement

dynamic resizing?

Investigate dynamic resizing of the data structure, rather than requiring the user to specify a capacity at initialization. I initially avoided this under the belief that using e.g. Vec's magic resizing behavior was risky when pushing the limits of my machine's memory, but I never verified that was true. Test on large datasets that are close to the memory limits of the machine it's running on. If it works, introduce new constructors that make capacity optional.

Decide what to do about stats tracked during search

They slow down stuff a little. Remove?

e.g., search_root_ancestory, search_depth.

Use correct error idioms

Define an error type and return Either where appropriate. Remove asserts.