embedcache

Use case

Batch RAG ingestion

Embed a large corpus once, re-run the pipeline as often as you like, and only pay CPU time for the chunks that actually changed.

The problem

RAG ingestion re-embeds the same documents on every pipeline run. Against a hosted API that means repeated per-token charges and rate-limited backfills that stretch into days.

How embedcache helps

  1. 1

    Point your ingestion job at the embedcache /v1/embed or /v1/process endpoint instead of a hosted embedder.

  2. 2

    Each chunk is keyed by its content hash and model, so unchanged chunks are cache reads on subsequent runs.

  3. 3

    Re-run the pipeline freely: only genuinely new or edited chunks trigger inference.

  4. 4

    Hand the resulting vectors to your vector database (Qdrant, pgvector, LanceDB) for indexing and search.

Features involved