README

Pure-PHP utilities for local-first AI pipelines.
Text chunkers, packed-vector math, Matryoshka truncation — boring on purpose.

What is ai-toolkit?

The plumbing every embed-and-search pipeline needs and no one wants to rewrite: split documents into chunks, move vectors around in the packed float32 format, slice Matryoshka embeddings down to size. Pure PHP, no extension required, no framework, no model downloads.

composer require displace/ai-toolkit

It pairs with ext-infer (embeddings), ext-turbovec (vector search), and the ai-contracts interfaces — but depends on none of them.

Chunkers

Three strategies behind one Chunker interface, sized in characters (multibyte-aware):

use Displace\AI\Toolkit\Text\RecursiveCharacterChunker;

$chunker = new RecursiveCharacterChunker(size: 1000, overlap: 200);

foreach ($chunker->chunk($markdown) as $chunk) {
    // embed → index
}

Packed-vector math

Vectors in the Displace stack travel as packed little-endian float32 binary strings — the output of pack('g*', ...$floats), batches by plain concatenation. Packed is the pure-PHP companion for glue code and tests:

use Displace\AI\Toolkit\Vector\Packed;

$a = Packed::pack([0.12, 0.48, /* ... */]);   // floats → packed buffer
$b = Packed::pack([0.33, 0.19, /* ... */]);

Packed::cosine($a, $b);                       // similarity in [-1, 1]
Packed::dot($a, $b);                          // inner product
Packed::norm($a);                             // L2 norm
Packed::normalize($batch, dim: 1024);         // per-vector unit length
Packed::unpack($a, dim: 1024);                // packed buffer → floats

Matryoshka (MRL) truncation

MRL-trained embedding models (Qwen3-Embedding, ...) pack the most important information into the leading coordinates, so a prefix of the vector is itself a usable embedding. Trade recall for a smaller index by slicing — no re-embedding required:

// 1024-dim Qwen3 embeddings down to 256-dim, renormalized per vector:
$small = Packed::truncate($vectors, fromDim: 1024, toDim: 256);

Only valid for MRL-trained models; truncating an ordinary embedding just throws information away.

On the hot path

Packed is for correctness, not speed — it's pure PHP. When ext-turbovec is loaded, prefer its native Displace\Vector\Vectors pack/unpack and let the index do the scoring; when you only need similarity between a handful of vectors, Packed is plenty.

Deliberately out of scope

Token-based chunking — token counts are model-specific and need a tokenizer; character budgets are model-independent and close enough for retrieval (rule of thumb: ~4 characters per English token) · embedding generation — that's ext-infer or your API client · vector storage and ANN search — that's ext-turbovec or your database · document loaders/parsers (PDF, HTML, ...) — bring your own text · an orchestration framework — chains and agents belong to the frameworks.

License

MIT © 2026 Eric Mann / Displace Technologies