Available providers
Meilisearch supports a wide range of embedding providers, each with different models, pricing, and strengths:| Provider | Models | Strengths | Guide |
|---|---|---|---|
| OpenAI | text-embedding-3-small, text-embedding-3-large | Straightforward setup, good general quality | Guide |
| Cohere | embed-v4.0, embed-english-v3.0, embed-multilingual-v3.0 | Latest v4 supports text and images, strong multilingual | Guide |
| Voyage AI | voyage-4, voyage-4-lite, voyage-4-large | High quality, flexible dimensions, domain-specific models | Guide |
| Jina | jina-embeddings-v4, jina-embeddings-v5-text-small/nano | v4 supports text, images, and PDFs, 32K context | Guide |
| Mistral | mistral-embed | Good for existing Mistral users | Guide |
| Google Gemini | gemini-embedding-001 | High dimensions (3072), Google ecosystem | Guide |
| Cloudflare | bge-small/base/large, embeddinggemma, qwen3 | Edge network, low latency, free tier | Guide |
| AWS Bedrock | Titan v2, Nova, Cohere Embed v4 on Bedrock | AWS ecosystem, multimodal options | Guide |
| HuggingFace (local) | Any compatible model | No API costs, full control | Guide |
| HuggingFace Inference | bge, MiniLM, mpnet, multilingual-e5, and more | Scalable open-source models, hundreds available | Guide |
Smaller models are often better
Bigger is not always better. In a hybrid search setup, Meilisearch combines keyword results with semantic results using its smart scoring system. Full-text search already handles exact matches very well, so the semantic side only needs to capture general meaning, not every nuance. This means a small, fast embedding model is often enough. The quality difference between a 384-dimension model and a 3072-dimension model is rarely worth the extra cost and latency, especially when the keyword side is already covering precise queries. Prioritize cheaper, faster models unless you have a specific reason to need more dimensions or higher embedding quality. Models liketext-embedding-3-small, voyage-4-lite, jina-embeddings-v5-text-nano, or embed-english-light-v3.0 are excellent starting points.
What to look for
Cost and rate limits
Embedding providers charge per token or per request. For large datasets, embedding costs add up during indexing. Consider:- Free tiers: Cloudflare Workers AI and local HuggingFace models have no per-request cost
- Rate limits: free-tier accounts on paid providers may slow down indexing significantly. Meilisearch handles retries automatically, but higher tiers index faster
- Re-indexing: Meilisearch caches embeddings and only re-generates them when document content changes, reducing ongoing costs
Dimensions
Lower-dimension models are faster to index, use less memory, and produce faster searches. Higher dimensions can capture more semantic nuance but with diminishing returns.| Dimensions | Trade-off |
|---|---|
| 384 | Fast, low memory, good for most use cases |
| 768-1024 | Balanced quality and performance |
| 1536-3072 | Higher quality, slower, more memory |
Domain specialization
Some providers offer models specialized for specific domains:- Legal, medical, financial: check if your provider has domain-specific models or fine-tuned variants
- Multilingual: if your content is not in English, choose a model with explicit multilingual support (Cohere’s multilingual models, Jina v3/v5, or multilingual BGE models)
- Code: some models are optimized for code search
Indexing speed
Embedding generation is the main bottleneck during indexing. Two factors affect speed:- API latency: cloud providers add network round-trip time per batch. Providers with edge networks (Cloudflare) or regional endpoints (Bedrock) can be faster
- Model size: larger models take longer to compute embeddings, even on the provider side
Maximize performance with composite embedders
If you need the best possible indexing speed and search latency, consider using a composite embedder. This lets you use different models for indexing and search:- Indexing: use a cloud provider (Cloudflare Workers AI, HuggingFace Inference Endpoints, or any REST API) to generate high-quality embeddings at scale without impacting your Meilisearch server
- Search: use a local HuggingFace model (like
BAAI/bge-small-en-v1.5) running inside Meilisearch for near-instant query embedding with zero API latency