feat: add Jina Embeddings v5 text-nano and text-small as alternative models

[hanxiao]

Summary

Add Jina Embeddings v5 as alternative embedding models, configurable via LlamaCppConfig.embedModel.

Two models available as GGUF from HuggingFace:

MMTEB scores vs model size. jina-v5-text models outperform models 2-16x their size. (source)

MTEB English v2 scores. v5-text-nano (239M) achieves 71.0, matching models with 2x+ parameters. (source)

Both v5 models use a different task prefix format (Query: / Document: ) compared to embeddinggemma's nomic-style format. The format functions auto-detect Jina v5 URIs and apply the correct prefix.

Changes

• src/llm.ts: Add JINA_V5_NANO_EMBED_MODEL and JINA_V5_SMALL_EMBED_MODEL constants, isJinaV5Model() utility, model-aware formatQueryForEmbedding() / formatDocForEmbedding()
• src/qmd.ts: Pass embedModelUri to format functions at all embed call sites
• src/store.ts: Pass embedModelUri to format functions at all embed/query call sites
• README.md: Document alternative models, add v5 prompt format section
• CHANGELOG.md: Add entry under [Unreleased]

Usage

import { LlamaCpp, JINA_V5_SMALL_EMBED_MODEL } from "./llm.js";

const llm = new LlamaCpp({ embedModel: JINA_V5_SMALL_EMBED_MODEL });

Not changed

Default model remains embeddinggemma-300M. This is additive only.

Paper: arXiv:2602.15547 | Blog | MTEB Leaderboard