Multilingual support: French (and 100+ languages) via Ollama BGE-M3

[Motokiyo]

Problem

MemPalace v3.0.0 uses ChromaDB's default embedding model (all-MiniLM-L6-v2), which is primarily trained on English text. When used with French content (or other non-English languages), search results are irrelevant — negative similarity scores, wrong documents returned.

Solution

We created a patch that replaces the default embedding with BGE-M3 via Ollama, supporting 100+ languages while keeping everything local (no cloud dependency).

Changes:

• New ollama_embedding.py — ChromaDB-compatible EmbeddingFunction using Ollama's /api/embed endpoint
• Patched miner.py and searcher.py to use BGE-M3 when Ollama is available, with graceful fallback to MiniLM-L6

Results on our French corpus (1500 files):

Query: "détection de chute reachy care"	Before (MiniLM)	After (BGE-M3)
Top result	store.js (JS in-app purchases — zero relevance)	PROJET_ARISTOTE_VISION.md (Reachy Care stack — perfect match)
Score	-0.428 (negative = no match)	+0.101 (positive = relevant)

Patch

Available at: https://github.com/Motokiyo/mempalace-multilingual

One-line install:

git clone https://github.com/Motokiyo/mempalace-multilingual && cd mempalace-multilingual && ./install.sh

Suggestion for upstream

Consider making the embedding function configurable in mempalace.yaml:

embedding:
  provider: ollama  # or "default" for MiniLM
  model: bge-m3
  base_url: http://localhost:11434

This would let users choose their embedding model without patching source files. Happy to submit a PR if you're interested.

— Alexandre Ferran / EIFFEL AI

[NickShtefan]

Hi @Motokiyo — your French numbers (-0.428 → +0.101) are exactly the same shape as the Russian numbers we saw (0.19–0.40 → 0.70–0.77). Same root cause, and the configurable-embedding-function approach from PR #442 is a superset of what your Ollama-BGE-M3 patch does — it lets you point at any sentence-transformers model (including Ollama via the ollama:<model> provider that #488 adds on top), with the model stamped into ChromaDB collection metadata so MCP and CLI can't desync.

If you want to A/B BAAI/bge-m3 vs intfloat/multilingual-e5-base on your French corpus, the mempalace re-mine --model <name> flow in #442 makes the swap reproducible. Both are 100+ lang, but e5-base has a 512-token window vs BGE-M3's 8192 — the win/loss split should be content-length-dependent, would be useful real-world data for the merge discussion.