What each vector store costs when nobody queries it
The models are billed as they are used - the embedding model while documents are ingested and queries are embedded, the model that answers while it answers. The vector store is billed all month, queried or not. With 100,000 chunks of 1024 dimensions (about 0.61 GB, counting 2 KB of text and metadata per chunk), at us-east-1 prices from the AWS Price List as of 2026-10-02:
| Vector store | Per month, idle | Why |
|---|---|---|
| OpenSearch Serverless | $350.41 ($175.21 for testing) | A Classic collection bills at least 2 OCUs (1 indexing, 1 search) around the clock; with standby replicas off, 1 OCU. The first collection pays it; later ones with the same KMS key share the OCUs. |
| OpenSearch managed cluster | $27.50 | The smallest domain: one t3.small.search node with gp3 storage (10 GB at least). One node has no replica - production domains run three or more. |
| Aurora PostgreSQL | $0.06 | Aurora Serverless v2 with 0 minimum ACUs, as Bedrock's quick create sets it (0-16): the instance pauses when idle (after 5 minutes by default) and only storage is billed. The first query after a pause waits about 15 seconds (30+ after a day) while it resumes; each ACU-hour while running costs $0.12. |
| S3 Vectors | $0.04 | Storage only - no compute while idle. Each sync also pays $0.20 per GB put, and each query per request and data processed. |
OpenSearch Serverless, the console's default, is the one that surprises: $0.24 per OCU-hour for at least two OCUs is about $350 a month before the first query. Its next generation scales to zero, but AWS says NextGen collections do not work with the Knowledge Bases Retrieve API yet, so a knowledge base still needs a Classic collection. S3 Vectors and a paused Aurora Serverless v2 cost little more than their storage; S3 Vectors is meant for infrequent queries, Aurora wakes up in about 15 seconds. Other Regions price differently, and ingestion, queries and the models are not included.
Vector dimensions per embedding model
| Model | Model ID | Dimensions | Binary |
|---|---|---|---|
| Titan Text Embeddings V2 | amazon.titan-embed-text-v2:0 | 1024, 512, 256 | Yes |
| Titan Embeddings G1 - Text | amazon.titan-embed-text-v1 | 1536 | - |
| Embed English | cohere.embed-english-v3 | 1024 | Yes |
| Embed Multilingual | cohere.embed-multilingual-v3 | 1024 | Yes |
| Embed v4 | cohere.embed-v4:0 | 1536, 1024, 512, 256 | - |
| Titan Multimodal Embeddings G1 | amazon.titan-embed-image-v1 | 1024 | - |
| Amazon Nova Multimodal Embeddings | amazon.nova-2-multimodal-embeddings-v1:0 | 1024 | - |
The index's dimension and the knowledge base's dimensions must match. An index's dimension is fixed when it is created, so another dimension means a new index. Binary vectors are 32 times smaller than float32 ones but only OpenSearch Serverless and OpenSearch managed clusters store them, with the hamming space type (Euclidean l2 for float vectors); S3 Vectors takes float32 only.
Knowledge base limits that cannot be raised
None of these quotas is adjustable in Service Quotas (as of 2026-10-02):
| File size (text content) | 50 MB |
| Image file size (JPEG, PNG) | 3.75 MB |
| Ingestion job size | 100 GB |
| Files added or updated per ingestion job | 5,000,000 |
| Files with a foundation model or BDA parser | 1,000 |
| .metadata.json file size | 10 KB |
| Data sources per knowledge base | 5 |
| Knowledge bases per account and Region | 100 |
| Ingestion jobs at once (per account / per knowledge base) | 5 / 1 |
| User query size | 1,000 characters |
| Custom metadata per vector in S3 Vectors | 1 KB, 35 keys |
A bigger corpus is split, not raised: up to 5 data sources per knowledge base, each with its own inclusion prefixes, synced one after another.
Why a sync fails or ignores files
- Formats: .txt, .md, .html, .doc/.docx, .csv, .xls/.xlsx and .pdf; text files up to 50 MB. Other files are skipped - "View warnings" in the data source's sync history lists them.
- Metadata files are named after their document plus .metadata.json (report.pdf → report.pdf.metadata.json), sit next to it, stay under 10 KB and hold strings, numbers, booleans or string lists - otherwise the sync reports "Ignored files due to invalid metadata attributes".
- The embedding model runs in the knowledge base's Region with its own model ID, and the service role may invoke that model there.
- The vector index has exactly the dimensions of the embedding configuration; a mismatch fails every document.
- The index uses the faiss engine. With nmslib, metadata filtering does not work - create a new index with faiss and a new knowledge base on it.
- Custom metadata fields used in filters are keyword fields (or text with a keyword subfield); otherwise filtering fails with "Rewrite first".
- The data access policy names the service role, and a private collection's network policy allows bedrock.amazonaws.com.
- One ingestion job runs per data source and knowledge base, five per account; StartIngestionJob takes one request per 10 seconds - wait for COMPLETE before the next sync.
Frequently asked questions
Is anything I enter sent anywhere?
No. The configuration is put together in your browser. The page only counts that it was used, with which vector store and which first problem.
Should I set up a vector store at all?
AWS now recommends a Bedrock Managed Knowledge Base, which keeps the vectors in a datastore Bedrock runs for you and takes a custom embedding model only with float32 and 1,024 dimensions. This tool is for the customer-managed kind, where you choose and pay for the store - and for Bedrock's quick create, which builds the same stores.
Why are Pinecone, MongoDB Atlas, Redis and Neptune Analytics missing?
They are supported too. Pinecone, MongoDB Atlas and Redis Enterprise Cloud are set up and priced outside AWS; a Neptune Analytics graph for GraphRAG gets its vector dimension when it is created and needs no index mapping. This tool covers the four AWS stores you build the index in yourself.
References
Prerequisites for using a vector store you created for a knowledge base
Supported models and Regions for Amazon Bedrock knowledge bases
Create a service role for Amazon Bedrock Knowledge Bases
Sync your data with your Amazon Bedrock knowledge base
Amazon Bedrock endpoints and quotas
Selecting a vector store for Amazon Bedrock Knowledge Bases