Trait Filter

Source

pub trait Filter {
    // Required method
    fn allows(&self, tag: u64) -> bool;

    // Provided methods
    fn exact(&self, _id: u64) -> bool { ... }
    fn narrowing(&self) -> bool { ... }
}

Expand description

What decides whether the scan bothers with a member.

A filtered vector search is a recall lottery when the filter runs after the search: ask for ten English passages, get the best forty by vector, find three of them are English, and the other seven English passages that were nearer never had a chance. The fix is to filter inside the scan, so that only members that can be answers are ranked at all, and that means the thing the filter reads has to sit next to the codes rather than behind a lookup into somebody else’s table.

So every member carries a u64 tag, given at insert, and a filter is a predicate on that tag. What the tag means is the caller’s business. A handful of low cardinality attributes pack into it exactly, one field each, and the filter is then exact. Anything wider goes through Signature, which is exact in the direction that matters: it never rejects a member that should have matched, so the caller’s real predicate over the answers still decides. A tag that is only a summary needs a second test somewhere, and the place for it is Filter::exact, which sees the member’s id and can go and read whatever the caller keyed by it. That runs only for members the tag let through that are also near enough to be ranked, which is why it is allowed to be the expensive one: an expression over a JSON string is fine there and would not be fine in the scan.

Required Methods§

Source

fn allows(&self, tag: u64) -> bool

Whether a member with this tag is worth ranking.

Provided Methods§

Source

fn exact(&self, _id: u64) -> bool

The second test, on the member’s id rather than on its tag.

Everything lets everything through by default, because for a filter whose tag says the whole truth there is nothing left to ask. Override it when the tag is a summary and the real predicate lives in a table of the caller’s, and keep the tag test as the cheap superset of it: a member the tag rejects never reaches here.

Source

fn narrowing(&self) -> bool

Whether this filter can turn members away.

A search that can be turned away has to be ready to look further than probe partitions, because the nearest ones may not hold k members that pass, and getting ready to costs something before the first partition is read: the probe order has to be built widen times longer than the search will use if it never widens. Saying no here is how a search that cannot be turned away avoids paying for the case that cannot happen to it. The default is yes, because a filter that answers wrongly here returns short answers rather than slow ones.