# Amazon Bedrock Knowledge Base

> Bedrock knowledge bases bill mostly through their backing vector store, which runs continuously whether queries arrive or not, plus embedding-model tokens at ingestion and retrieval charges per query. ZopNight discovers knowledge bases on the 6-hour cycle, tracks knowledge-base and vector-store cost from Cost Explorer or CUR 2.0, and flags unqueried ones.

Source: https://zop.dev/integrations/aws/bedrock-knowledge-base
Updated: 2026-08-19

---

A Bedrock knowledge base connects foundation models to your data for retrieval-augmented generation, incurring vector store, embedding, and query charges. The backing vector database often bills continuously even when queries stop.

## The vector store is the bill

A knowledge base's own meters are modest: embedding-model tokens when documents are ingested and re-synced, and retrieval charges when queries run. The dominant cost usually lives one layer down, in the vector store the knowledge base requires. That is most commonly OpenSearch Serverless, whose compute units bill hourly around the clock regardless of query traffic, or a provisioned Aurora or third-party vector database with its own standing rates. RAG's dirty secret is that the retrieval layer prices like infrastructure while the demos price it like an API call: the knowledge base that answered three questions this month paid its vector store for every hour of it.

## Retrieval infrastructure, audited

ZopNight discovers knowledge bases through its Bedrock provider on the 6-hour cycle, tracking knowledge-base and vector-store cost together from Cost Explorer or CUR 2.0. Joining the two is the point: the unused-knowledge-base recommendation exists because the expensive half is invisible from the Bedrock side. A knowledge base whose retrieval calls have stopped looks free in isolation while its OpenSearch Serverless collection bills continuously; attribution connects the quiet API to the standing charge and makes the pair a single deletable finding.

## RAG pilots and their remainders

The lifecycle is familiar: a RAG proof of concept ingests a document corpus, impresses a demo audience, and stalls in productization, leaving the vector store billing hourly behind a knowledge base nobody queries. Sync schedules keep re-embedding refreshed sources for the same absent audience. And duplicate knowledge bases accumulate per team and per experiment, each dragging its own collection, where one shared index would carry the corpus once.

## Knowledge bases and their stores

The Bedrock console's Knowledge bases view lists each one with its data sources and vector store configuration. The audit follows the pointer: for every knowledge base, find the backing collection and its hourly cost, then ask when retrieval last ran. The answer sorts keepers from remainders quickly.