Thunderboltdy10/ixos-documents

Ixos Documents

Derived from minishlab/potion-retrieval-32M, made by MinishLab and released under the MIT license. Pinned upstream revision: 6fc8051fab2a1e0ee76689cf08c853792ac285e7.

The 512-dimensional vocabulary rows are stored as symmetric int8 values. For each row, the scale is its maximum absolute value divided by 127 (or 1 for an all-zero row). Values are rounded to the nearest integer with ties to even, and clipped to [-127, 127]. Any upstream per-token weight multiplies the row scale. Int8 rows are stored as embeddings, and float32 scales as weights, in safetensors format. Tokenizer and configuration files are unchanged.

No training or distillation was done. The encoded Rust loader decodes the used rows and pools in float32 on CPU; this is a storage transformation, not integer inference. Run python scripts/models.py build to reproduce the files and verify the hashes against the Rust installer table. The publication includes UPSTREAM_MODEL_CARD.md with the original authorship, license and training details.

Use with Ixos

This is a runtime artifact for Ixos, a local CPU search engine. Install it with:

ixos model install documents

Ixos downloads the three runtime files at a pinned revision, verifies their SHA-256 hashes and loads them with its Rust loader. The model files are stored outside the executable and reused offline. This repository does not provide a general-purpose Transformers pipeline.