potion-retrieval-32M-obsidian-tfjs

potion-retrieval-32M-obsidian-tfjs

Fine-tuned variant of minishlab/potion-retrieval-32M, adapted for Obsidian-markdown content (headings/breadcrumbs, wikilinks, tags, frontmatter-style metadata, code fences) and packaged for transformers.js / Smart Connections.

• Training: MNRL contrastive (lr 8e-2, 3 epochs) on 13,112 (breadcrumb/title/tags anchor, note-block) pairs mined from personal medical study vaults
• Vault holdout retrieval (453 held-out files): R@5 0.439 → 0.662, R@10 0.467 → 0.749 vs base
• General-domain guard: STS-B spearman 0.8285 → 0.8249

Same interface as the base tfjs export: input_ids [batch, seq] + attention_mask [batch, seq] (int64) → last_hidden_state [batch, 1, 512] masked mean-pooled; tokenizer ships without special-token post-processor (matches native model2vec pooling).

import { pipeline } from '@huggingface/transformers';
const p = await pipeline('feature-extraction', 'mikhailmedical/potion-retrieval-32M-obsidian-tfjs', { dtype: 'fp32' });
const out = await p(texts, { pooling: 'mean', normalize: true });

MIT license, inherited from the upstream model.