Search System Upgrade: Faster, Local, and More Reliable

What Happened

Over the past week, some users may have experienced issues with SeferAI's text search functionality. We want to explain what happened and what we've done about it.

Google recently deprecated their `text-embedding-004` API, which SeferAI relied on to power semantic search across our library of Jewish texts. When the API was retired, search queries began returning errors instead of results.

What We Did

Rather than migrating to another external API, we moved to a fully local embedding model running directly on our servers. Here's what changed:

• New model: We now use `all-MiniLM-L6-v2`, a proven open-source sentence transformer
• Re-indexed everything: All 156,185 texts in our collection have been re-embedded and re-indexed
• Zero external dependencies: Search no longer relies on any third-party API -- it runs entirely on our infrastructure

What This Means for You

• Faster search: Query embedding now takes ~50ms locally vs ~250ms with the old external API
• More reliable: No more dependency on external services that can be deprecated without notice
• Same quality: Search relevance remains excellent across Torah, Talmud, Midrash, and all other collections
• Zero added cost: The local model runs efficiently within our existing server resources

Looking Ahead

This incident reinforced our commitment to self-hosted, resilient infrastructure. By keeping critical components local, we ensure SeferAI remains available and responsive regardless of changes to third-party services.

All systems are fully operational. If you notice any issues, please don't hesitate to reach out.

-- The SeferAI Team