Knowledge store (vector store): what it is and how it is attacked
The collection of documents an AI assistant searches before it answers.
Also known as
- vector store
- vector database
- embeddings store
- RAG store
- retrieval store
- knowledge base
- index
Typing any of them into the editor finds this object.
Why it matters on a security diagram
It decides what the assistant knows, which makes it both a disclosure surface and an injection surface. Anything that can write into it can change what every answer is based on, and often nobody is reviewing what goes in.
How it gets attacked, and what reduces it
How it gets attacked
- Documents added by anyone, including from a shared mailbox or an open folder, so writing into it needs no access to the model at all
- Contents retrieved for a person who would not be allowed to open the source document, because the permissions did not travel with the text
- Hostile text stored once and returned in answers long afterwards
What reduces it
- Carry the source document's permissions through to retrieval, so the assistant cannot answer from something the reader could not open
- Control what may be written into it as carefully as what may be read out
- Keep the source of every passage, so an answer can be traced back and a poisoned document can be found and removed
- Re-check the store after any incident involving a system that could write to it
Where it sits
- Group
- Cloud and subscribed services · Things that only exist because somebody else runs the platform underneath them.
- Whose side, by default
- Ours · Belongs to the organisation the diagram is about.
- Catalogue identifier
- retrieval-store
Reviewed . CC BY 4.0.
Others in cloud and subscribed services
- Cloud platformAWS · Azure · GCP · Google Cloud · public cloud · hyperscaler · IaaSA supplier's computing platform, rented by the hour, where a setting made on a website creates or removes a whole system.
- AI model endpointLLM API · inference endpoint · model service · AI serviceWhere an application sends a question, along with documents, images or data, to an AI model and gets an answer back.
- Cloud accountsubscription · project · AWS account · Azure subscription · Google Cloud project · cloud environmentOne named compartment of a cloud platform, with its own bill and its own list of who may change what.
- Cloud account groupmanagement group · AWS Organizations · GCP folder · landing zone · account structureThe parent that owns all of an organisation's cloud accounts and can set rules over every one of them.
- Cloud control panelcontrol plane · management console · cloud portal · cloud API · management planeThe website and commands used to configure the cloud itself, as opposed to the systems running inside it.
- Supplier-run applicationSaaS · software as a service · cloud application · business applicationAn application a supplier runs entirely, which the organisation subscribes to and puts its data into.
- Connected appOAuth application · app registration · enterprise application · app consent · add-inAn outside program that somebody has allowed to reach company data on their behalf.
- Managed databaseRDS · Azure SQL · Cloud SQL · database as a service · DBaaSA database the cloud provider runs, where the data and the settings are yours but the machine is never touched.
- On-demand functionserverless · Lambda · Azure Functions · Cloud Run · FaaSA small piece of code that runs only when something asks for it, with no server to look after.
- Container platformKubernetes · K8s · cluster · EKS · AKS · GKE · OpenShiftThe system that decides where each packaged application runs, and starts it again when it stops.
- Container image storecontainer registry · image registry · ECR · ACR · artefact storeThe library of packaged applications that machines fetch from whenever they start something.
- Infrastructure blueprintinfrastructure as code · IaC · Terraform · state file · CloudFormation · BicepThe written description of what the cloud should contain, together with the record of what was actually built.
- Source code repositoryrepo · repository · GitHub · GitLab · Bitbucket · Azure DevOps · source control · version control · gitWhere the organisation's own code and configuration is kept and changed.
The cloud and subscribed services group lists all 14 of them side by side.
Knowledge store on your own diagram
Open the editor, press N, and type vector store. The object is placed and connected to whatever was selected, and Tab adds the next one already joined to it. Nothing is uploaded: the page is served with a Content Security Policy that forbids the browser from making any outbound request at all.
Open the diagram maker