Skip to content

[FEATURE] Add larger Chinese/multilingual embedding models to FastEmbed provider (e.g. bge-base-zh-v1.5, bge-m3) #1295

Description

@bannana

Title: Add larger Chinese/multilingual embedding models to FastEmbed provider (e.g. bge-base-zh-v1.5, bge-m3)

Environment: basic-memory 0.22.1, fastembed 0.8.0

Problem
The local fastembed provider bundles exactly one Chinese model (bge-small-zh-v1.5, 24M params). For Chinese notes this small model yields weak embeddings — cosine similarities cluster in the 0.3–0.5 range and the default semantic_min_similarity: 0.55 filters most results out. There is no local option to use a larger Chinese or multilingual model.

Request
Add larger CJK-capable models to the bundled FastEmbed list, e.g.:

  • BAAI/bge-base-zh-v1.5 (102M, 768-dim)
  • BAAI/bge-large-zh-v1.5 (324M, 1024-dim)
  • BAAI/bge-m3 (multilingual, 568M, 1024-dim)

Why
bge-small-zh-v1.5 is the only CJK option and its small size caps semantic-retrieval quality for Chinese users. Users with capable hardware (e.g. Apple Silicon) would prefer a larger model.

Workaround
None clean locally. Requires switching semantic_embedding_provider to openai (cloud API) or litellm + a local model server (e.g. Ollama), both of which add external dependencies and undo the local-first design.

Metadata

Metadata

Assignees

No one assigned

    Labels

    enhancementNew feature or request

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions