Summary

• document how to run GBrain embeddings against a local Ollama server via the OpenAI-compatible /v1/embeddings endpoint
• make the embedding model and dimensions configurable via environment variables so local models can be used without editing source each time
• add verification notes about the 1536-dimension schema compatibility caveat

Why

I validated this locally with qwen3-embedding:4b on Ollama. The important detail is that Ollama's native /api/embed can return a model's native dimension, while the OpenAI-compatible endpoint can return 1536 dimensions when requested, which matches GBrain's current schema.

Test Plan

• bun test test/embed.test.ts
• verified locally that gbrain embed --stale and gbrain query work against a local Ollama server configured through OPENAI_BASE_URL, OPENAI_EMBED_MODEL, and OPENAI_EMBED_DIMENSIONS