docs: remove stale distanceSQ8 JSDoc left by the SQ8 hook refactor
The native SQ8 distance hook renamed the original distanceSQ8 to distanceSQ8Js and added a dispatcher in its place, but left the old function's doc block orphaned above distanceSQ8Js (two consecutive comments, the first dangling). Merge them into one accurate block on distanceSQ8Js with dash-notation params.
This commit is contained in:
parent
4b6f63ed67
commit
6099101336
1 changed files with 16 additions and 23 deletions
|
|
@ -96,31 +96,24 @@ export function dequantizeSQ8(quantized: Uint8Array, min: number, max: number):
|
||||||
}
|
}
|
||||||
|
|
||||||
/**
|
/**
|
||||||
* Compute approximate cosine distance between two SQ8-quantized vectors.
|
* Reference JS implementation of {@link distanceSQ8}: approximate cosine distance
|
||||||
|
* between two SQ8-quantized vectors, operating directly on the uint8 arrays without
|
||||||
|
* full dequantization. Cosine distance is `1 - (A·B / (|A| · |B|))`; for quantized
|
||||||
|
* values `q_i` with codebook `(min, max)` the actual value is
|
||||||
|
* `min + q_i * (max - min) / 255`, and the dot product / norms are computed on the
|
||||||
|
* quantized values and rescaled by the codebook parameters.
|
||||||
*
|
*
|
||||||
* Operates directly on uint8 arrays without full dequantization.
|
* Exported so callers can explicitly restore the default after swapping in a native
|
||||||
* Uses integer arithmetic where possible for speed.
|
* implementation, and so cross-language parity tests can compare native output
|
||||||
|
* against this baseline.
|
||||||
*
|
*
|
||||||
* Cosine distance = 1 - (A.B / (|A| * |B|))
|
* @param a - First quantized vector.
|
||||||
*
|
* @param aMin - First vector's codebook minimum.
|
||||||
* For quantized values q_i with codebook (min, max):
|
* @param aMax - First vector's codebook maximum.
|
||||||
* actual_i = min + q_i * (max - min) / 255
|
* @param b - Second quantized vector.
|
||||||
*
|
* @param bMin - Second vector's codebook minimum.
|
||||||
* The dot product and norms can be computed on quantized values
|
* @param bMax - Second vector's codebook maximum.
|
||||||
* and scaled by the codebook parameters.
|
* @returns Approximate cosine distance in [0, 2].
|
||||||
*
|
|
||||||
* @param a - First quantized vector
|
|
||||||
* @param aMin - First vector codebook minimum
|
|
||||||
* @param aMax - First vector codebook maximum
|
|
||||||
* @param b - Second quantized vector
|
|
||||||
* @param bMin - Second vector codebook minimum
|
|
||||||
* @param bMax - Second vector codebook maximum
|
|
||||||
* @returns Approximate cosine distance [0, 2]
|
|
||||||
*/
|
|
||||||
/**
|
|
||||||
* Reference JS implementation of {@link distanceSQ8}. Exported so callers can
|
|
||||||
* explicitly restore the default after swapping in a native implementation, and
|
|
||||||
* so cross-language parity tests can compare native output against this baseline.
|
|
||||||
*/
|
*/
|
||||||
export function distanceSQ8Js(
|
export function distanceSQ8Js(
|
||||||
a: Uint8Array,
|
a: Uint8Array,
|
||||||
|
|
|
||||||
Loading…
Add table
Add a link
Reference in a new issue