Embedding Models
zebraindex ships with a registry of 13 embedding models. Models are downloaded automatically on
first use from HuggingFace. Switch models with --model:
zebraindex --model BAAI/bge-small-en-v1.5 index .Note: changing the model requires re-indexing — embeddings from different models are not compatible.
Code-optimised (recommended for codebases)
| Model | Params | Context | Notes |
|---|---|---|---|
jinaai/jina-embeddings-v2-base-code | 137M | 8192 | Default. State-of-the-art for code search. 30+ languages. |
General purpose — small (CPU-friendly)
| Model | Params | Notes |
|---|---|---|
sentence-transformers/all-MiniLM-L6-v2 | 22.7M | Fastest; minimal memory |
sentence-transformers/all-MiniLM-L12-v2 | 33.4M | Slightly better than L6 |
BAAI/bge-small-en-v1.5 | 33.4M | Best accuracy in this size class |
intfloat/e5-small-v2 | 33.4M | Requires query:/passage: prefixes |
thenlper/gte-small | 33.4M | No prefixes needed; robust on varied text |
General purpose — base (GPU recommended)
| Model | Params | Notes |
|---|---|---|
BAAI/bge-base-en-v1.5 | 109M | Excellent accuracy/speed balance |
intfloat/e5-base-v2 | 109M | High accuracy; strict prefix routing |
thenlper/gte-base | 109M | No prefixes; competes with BGE/E5 |
Multilingual
| Model | Params | Languages | Notes |
|---|---|---|---|
intfloat/multilingual-e5-small | 118M | 100+ | Good for mixed-language repos |
intfloat/multilingual-e5-base | 278M | 100+ | GPU recommended |
BAAI/bge-m3 | 567M | 100+ | Heavy; only dense retrieval in this mode |
Hardware guide
| Hardware | Recommended models |
|---|---|
| CPU only | MiniLM-L6, bge-small, gte-small |
| Metal (macOS M-series) | jina-code (default), bge-base, gte-base |
| CUDA GPU (≥4 GB VRAM) | Any model up to bge-base |
| CUDA GPU (≥12 GB VRAM) | bge-m3, multilingual-e5-base |
Run zebraindex doctor to see which device zebraindex is using and whether the current model
fits in available memory.
Prefix routing
Models from the E5 and Jina families use instruction prefixes at inference time:
- Query prefix — prepended when embedding a search query (e.g.
query:) - Passage prefix — prepended when embedding a chunk at index time (e.g.
passage:)
Configure custom prefixes when starting the daemon:
zebraindex --query-prefix "query: " --passage-prefix "passage: " index .BGE, GTE, and MiniLM models do not require prefixes.