Global Architecture

Libro fundamentally rewrites how vectorization occurs by bringing the embedding models to the edge.

The Problem with Traditional Vectors

Historically, developers had to:

1. Send text to OpenAI to get an embedding.
2. Store that massive float array in Pinecone or Postgres (pgvector).
3. Do this on a centralized database, introducing heavy latency.

The Libro Edge Engine

Libro runs specialized, highly optimized ONNX variants of Transformer models directly on the Edge. When you call ingest():

1. The text hits our nearest edge node (e.g. Frankfurt, Sydney, Washington).
2. It is chunked and vectorized locally in under 15ms.
3. It is stored in a distributed edge-replica database.

This architecture results in zero central-database bottleneck, meaning infinite scaling.