---
license: mit
base_model: intfloat/multilingual-e5-base
library_name: transformers.js
pipeline_tag: feature-extraction
tags: [sentence-similarity, onnx, vocabulary-trimmed, e5]
---

# multilingual-e5-base-iconclass

Vocabulary-trimmed ONNX (int8) copy of [`Xenova/multilingual-e5-base`](https://huggingface.co/Xenova/multilingual-e5-base) (from [`intfloat/multilingual-e5-base`](https://huggingface.co/intfloat/multilingual-e5-base)), used as the **query encoder** of the Rijksmuseum Iconclass MCP server.

The XLM-R tokenizer's 250,002 pieces were reduced to **116,522** by dropping rows of the word-embedding table and the matching tokenizer entries. All other weights are unchanged. Kept pieces:

- every piece used by the server's corpus (Iconclass texts and keywords in all 13 languages);
- every single-character piece, so no input maps to `<unk>`;
- pieces of the 30,000 most frequent words (via [wordfreq](https://github.com/rspeer/wordfreq)) in the 13 Iconclass languages (English, German, French, Italian, Portuguese, Finnish, Dutch, Spanish, Polish, Czech, Hungarian, Japanese, Chinese), and the 3,000 most frequent in 13 further languages (incl. Russian, Greek, Arabic, Hebrew, Korean, Scandinavian).

Unigram tokenization picks the highest-scoring segmentation, so any text whose original segmentation uses only kept pieces tokenizes and embeds **identically** to the base model; vectors produced with the base model remain compatible. Other text is split into smaller kept pieces, giving a close but not identical vector.

Loaded with transformers.js (`dtype: "q8"`), resident memory drops from about 1,330 to 776 MB.

Built with [`scripts/trim-embedding-vocab.py`](https://github.com/kintopp/rijksmuseum-mcp-plus/blob/main/scripts/trim-embedding-vocab.py) (in the Rijksmuseum MCP+ repo); `trim_manifest.json` records the kept ids and inputs. Use the e5 prefixes (`query: ` / `passage: `) as with the base model.