Providers¶

Seamless-RAG uses pluggable providers for both embedding and LLM operations. Switch providers via environment variables or .env file.

Embedding Providers¶

SentenceTransformers (default)¶

Local inference, no API key required. Uses all-MiniLM-L6-v2 (384 dimensions).

EMBEDDING_PROVIDER=sentence-transformers
EMBEDDING_MODEL=all-MiniLM-L6-v2

Best for: quick start, offline use, CI/testing.

Gemini¶

Google's embedding API via the google-genai SDK. Supports MRL dimension control.

EMBEDDING_PROVIDER=gemini
EMBEDDING_MODEL=gemini-embedding-001
EMBEDDING_API_KEY=your-gemini-api-key
EMBEDDING_DIMENSIONS=768

Best for: production quality, free tier available.

OpenAI¶

OpenAI's embedding API. Highest dimensionality.

EMBEDDING_PROVIDER=openai
EMBEDDING_MODEL=text-embedding-3-large
EMBEDDING_API_KEY=your-openai-api-key
EMBEDDING_DIMENSIONS=3072

Ollama¶

Local embedding via Ollama REST API. No API key needed.

EMBEDDING_PROVIDER=ollama
EMBEDDING_MODEL=nomic-embed-text
LLM_BASE_URL=http://localhost:11434
EMBEDDING_DIMENSIONS=768

Foreign Model Auto-Correction¶

When switching providers, model names are automatically corrected. If you switch from gemini to sentence-transformers but forget to change the model name, the factory detects the mismatch and uses the correct default.

LLM Providers¶

Ollama (default)¶

Local LLM via Ollama REST API. No API key required.

LLM_PROVIDER=ollama
LLM_MODEL=qwen3:8b
LLM_BASE_URL=http://localhost:11434

Gemini¶

LLM_PROVIDER=gemini
LLM_MODEL=gemini-2.5-flash
LLM_API_KEY=your-gemini-api-key

OpenAI¶

LLM_PROVIDER=openai
LLM_MODEL=gpt-4o
LLM_API_KEY=your-openai-api-key

Custom Providers¶

Both layers use typing.Protocol. Implement the interface and it works without inheritance:

Embedding Protocol¶

class MyEmbeddingProvider:
    @property
    def dimensions(self) -> int:
        return 1024

    def embed(self, text: str) -> list[float]:
        return my_api.embed(text)

    def embed_batch(self, texts: list[str], batch_size: int = 32) -> list[list[float]]:
        return [self.embed(t) for t in texts]

LLM Protocol¶

class MyLLMProvider:
    def generate(self, prompt: str, context: str) -> str:
        return my_api.generate(prompt=prompt, context=context)

Register your provider in the factory module or pass it directly to SeamlessRAG.

Recommended Configurations¶

Use Case	Embedding	LLM	Notes
Quick start / offline	sentence-transformers	ollama	No API keys needed
Best quality	gemini	gemini	Free tier, strong models
Enterprise	openai	openai	Highest dimensions
Mixed	gemini (embed)	ollama (LLM)	Cloud embeddings, local inference