Repository navigation
[FEATURE] Add larger Chinese/multilingual embedding models to FastEmbed provider (e.g. bge-base-zh-v1.5, bge-m3) #1295
Description
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.
Activity
phernandez commentedon Aug 23, 2026
Basic Memory already passes
semantic_embedding_modelthrough to FastEmbed, so users can choose any text embedding model registered by their installed FastEmbed version. The model's actual output size must also be set throughsemantic_embedding_dimensions, followed bybm reindex --embeddings.The exact BGE variants requested here are not in FastEmbed 0.8.0's text-embedding catalog, so adding those model artifacts, pooling rules, and metadata would belong upstream rather than in a Basic Memory allowlist. The same catalog already includes larger local Chinese/multilingual choices, including
jinaai/jina-embeddings-v2-base-zhat 768 dimensions.#1300 documents how to list the installed FastEmbed catalog, configure a supported model and dimensions, and rebuild embeddings. We're treating this as a documentation/discoverability issue rather than adding model-specific automation.