| Overview |  |
Authors: Steve Roques.
em_filter_vec — Capability list → semantic f32 vector
Converts an agent's capability list into a unit-norm f32 binary vector suitable for cosine-similarity routing in em_pop.
1. Each capability string is normalised to a binary (atoms and
char-lists are accepted for convenience).
2. It is hashed via erlang:phash2/2 to a slot index in
[0, Dim-1]. phash2 is deterministic across BEAM restarts
for the same Erlang major version, so the same capability
always maps to the same slot.
3. The slot's weight is incremented by 1.0 (additive: two
capabilities that collide in the same slot reinforce each
other rather than cancelling).
4. The resulting weight vector is L2-normalised to unit length
so that cosine similarity is always well-defined and lies in
[-1, 1].
A uniform unit vector is returned (all slots equal, total norm = 1). This avoids a zero vector which has no well-defined cosine.
Dim to from_capabilities/2
when finer resolution is needed (e.g. 128 or 256).
| from_capabilities/1 | Derive a 64-dimension unit-norm f32 vector from a capability list. |
| from_capabilities/2 | Derive a Dim-dimension unit-norm f32 vector from a capability list. |
from_capabilities(Caps::[binary() | atom() | string()]) -> binary()
Derive a 64-dimension unit-norm f32 vector from a capability list.
Equivalent tofrom_capabilities(Caps, 64).
from_capabilities(Caps::[binary() | atom() | string()], Dim::pos_integer()) -> binary()
Derive a Dim-dimension unit-norm f32 vector from a capability list.
The output is a flat binary of Dim IEEE-754 single-precision floats
in little-endian byte order, as expected by kvex.
Vec = em_filter_vec:from_capabilities([<<"rss">>, <<"search">>], 64).
%% Vec is a 256-byte binary, unit norm, usable directly with kvex.
Generated by EDoc