Local Ollama

A good fit for MacBooks, Mac minis, desktops, capable NAS devices, or VPS boxes. Start with bge-m3 and dimension 1024; on Ollama it is about 567M parameters and works well for multilingual RSS.

If you want a lighter local setup, nomic-embed-text with dimension 768 is the simpler tradeoff. Local embedding keeps vector generation inside your own machine.

SiliconFlow / Gemini

For China-friendly OpenAI-compatible access, SiliconFlow BAAI/bge-m3 is free, uses dimension 1024, and is rate-limited by RPM / TPM. Gemini users can start with gemini-embedding-001, dimension 768, which fits a small personal RSS habit.

The provider creates vectors; ranking, profile, explanations, and index state remain inside your instance. Failures are captured as diagnostics and graceful fallback, not as a broken reader.

Index maintenance

Active indexes can backfill missing or stale embeddings. sqlite-vec can rebuild from the local authority table without calling the provider again.

Diagnostics expose coverage, candidate size, pending/failed jobs, clusters, ranked inventory, and warnings without leaking API keys or raw vectors.