Supported Embedding Models

PRX-Memory supports three embedding provider families. Each provider connects through the prx-memory-embed crate's unified adapter interface.

OpenAI-Compatible

Any API that follows the OpenAI embedding endpoint format (/v1/embeddings) can be used. This includes OpenAI itself, Azure OpenAI, and local inference servers.

PRX_EMBED_PROVIDER=openai-compatible
PRX_EMBED_API_KEY=your_openai_key
PRX_EMBED_MODEL=text-embedding-3-small
PRX_EMBED_BASE_URL=https://api.openai.com  # optional

Model	Dimensions	Notes
text-embedding-3-small	1536	Good balance of quality and cost
text-embedding-3-large	3072	Highest quality, higher cost
text-embedding-ada-002	1536	Legacy model

Local Inference

For privacy-sensitive deployments, point PRX_EMBED_BASE_URL to a local inference server running an open-source embedding model (e.g., via Ollama, vLLM, or text-embeddings-inference).

Jina AI

Jina provides high-quality multilingual embedding models optimized for retrieval tasks.

PRX_EMBED_PROVIDER=jina
PRX_EMBED_API_KEY=your_jina_key
PRX_EMBED_MODEL=jina-embeddings-v3

Model	Dimensions	Notes
jina-embeddings-v3	1024	Latest multilingual model
jina-embeddings-v2-base-en	768	English-optimized
jina-embeddings-v2-base-code	768	Code-optimized

Fallback Key

If PRX_EMBED_API_KEY is not set, the system checks JINA_API_KEY as a fallback.

Google Gemini

Gemini embedding models are available through the Google AI API.

PRX_EMBED_PROVIDER=gemini
PRX_EMBED_API_KEY=your_gemini_key
PRX_EMBED_MODEL=text-embedding-004

Model	Dimensions	Notes
text-embedding-004	768	Current recommended model
embedding-001	768	Legacy model

Fallback Key

If PRX_EMBED_API_KEY is not set, the system checks GEMINI_API_KEY as a fallback.

Choosing a Model

Priority	Recommended Model	Provider
Best quality	text-embedding-3-large	OpenAI-compatible
Best for code	jina-embeddings-v2-base-code	Jina
Multilingual	jina-embeddings-v3	Jina
Privacy / local	Any local model via openai-compatible	Self-hosted
Cost-effective	text-embedding-3-small	OpenAI-compatible

Switching Models

When switching embedding models, existing vectors become incompatible with the new model's vector space. Use the memory_reembed tool to re-embed all stored memories with the new model:

{
  "jsonrpc": "2.0",
  "id": 1,
  "method": "tools/call",
  "params": {
    "name": "memory_reembed",
    "arguments": {}
  }
}

WARNING

Re-embedding requires API calls for every stored memory. For large databases, this may take significant time and incur API costs. Plan re-embedding during low-usage periods.

Next Steps

• Batch Processing -- Efficient bulk embedding
• Reranking Models -- Second-stage reranking model options
• Configuration Reference -- All environment variables