Feature request: Pluggable embedding providers (Gemini, Voyage, Ollama)

[thephilipjohnson]

Problem

gbrain v0.18.2 hard-codes OpenAI's text-embedding-3-large in src/core/embedding.ts. Builders who don't use OpenAI as a vendor are forced to either add OpenAI just for embeddings or run lexical-search-only mode (semantic similarity ranking and hybrid query/ask degrade to keyword matching).

This pushes builders toward vendor sprawl. Many of us run on Gemini (Google), Anthropic (Claude), or local-first stacks deliberately.

Proposal

Abstract the embedding provider behind an interface and add provider switching:

• Default: OpenAI text-embedding-3-large (1536 dims) — current behavior, no breaking change
• Add: Gemini text-embedding-004 (768 dims, free tier on Google AI Studio, generous limits)
• Add: Voyage voyage-3-lite (512 dims, Anthropic's recommended embedding partner, free tier)
• Add: Ollama nomic-embed-text (768 dims, fully local, $0 cost, privacy-preserving)

Selected via:

• gbrain init --embedder gemini
• gbrain config set embedder gemini
• Or env: GBRAIN_EMBEDDER=gemini

Schema change: store embedding model + dimensions per chunk. Allows mixed-dimension corpora during migration. Migration helper: gbrain reindex --embedder gemini --all.

Why this matters

• Removes vendor lock to OpenAI for builders who chose other AI stacks deliberately
• Lowers the barrier to entry for users who don't have OpenAI billing
• Aligns with the broader gstack ethos of "use what you have"
• Local Ollama option preserves privacy for sensitive corpora

Willing to contribute

Happy to discuss design before submitting a PR. The lift is roughly:

• Refactor src/core/embedding.ts to an EmbeddingProvider interface
• Add 3 implementations (Gemini, Voyage, Ollama)
• Schema migration to track model + dims per chunk
• Update gbrain init flow to ask which embedder
• Tests against fixture corpus per provider

Estimated ~150-200 LOC plus migration. Open to alternative designs.

[blockfeed]

+1 on this.

Currently gbrain's embedding service only supports OpenAI API, which forces agents to depend on external paid APIs even when they already have local embedding infrastructure running (llama/llama-swap, Ollama, etc.).

Adding support for configurable embedding providers would enable:

Local/embedded inference for cost reduction
Reduced latency for local deployments
Vendor lock-in avoidance