counterparts-model-potion

The potion-base-8M static embedding table, packaged so counterparts can give its memory a semantic channel without an API key and without the network.

It is data only: two model files, this README, a license. No code, no install scripts, no dependencies.

What is in it

| file | what | sha256 | |---|---|---| | model.safetensors | one tensor, embeddings, F32, shape [29528, 256] (30,236,760 bytes) | f65d0f325faadc1e121c319e2faa41170d3fa07d8c89abd48ca5358d9a223de2 | | vocab.txt | the WordPiece vocabulary, one token per line, line number = row (29,528 lines) | 1394523a67ddd404a825428018c0582a6998bcfa044ecbcbf1f4d71adb94c61c |

Both files are copied byte-for-byte from https://huggingface.co/minishlab/potion-base-8M at revision bf8b056651a2c21b8d2565580b8569da283cab23 (last modified upstream 2026-03-27). The sha256 of model.safetensors is the one Hugging Face's LFS pointer states for that revision. Check them yourself:

shasum -a 256 model.safetensors vocab.txt

potion-base-8M is by MinishLab, MIT licensed: a Model2Vec model distilled from BAAI/bge-base-en-v1.5 and trained with Tokenlearn. See LICENSE for the full attribution.

How counterparts finds it

counterparts' static embedder (src/core/embed/static.ts) looks for a weights directory in this order:

1. an explicit directory passed by the caller;
2. the COUNTERPARTS_STATIC_WEIGHTS_DIR environment variable;
3. this package, resolved from counterparts' own location (require.resolve("counterparts-model-potion/package.json")), which is how an ordinary npm install counterparts finds it once counterparts lists it as a dependency.

The identity comes from the bytes, not from this package. At every load counterparts computes the sha256 of both files; if this package.json declares them (counterparts.sha256) they must match, or the table is refused by name (HASH_MISMATCH). The pair of hashes is what names the table in a store's tag (potion-base-8M@256, from counterparts' own list of known tables) — so the same two files give the same identity however they are found, and a changed table can never pass for this one. counterparts.model is a display label only.

How a text becomes a vector

The HF BERT basic tokenizer (clean, CJK padding, accent strip, lowercase, split on whitespace and punctuation), greedy WordPiece with ## continuations, drop [UNK], look up each id's row, take the mean, L2-normalize. 256 numbers, about 0.03 ms per text on an Apple M3 Pro, matching model2vec's own encode().