Vector stores

Create a vector store, load documents into it from an automation, search it, and let agents use it as a knowledge tool.

A vector store holds text documents with their embeddings, so an agent or an automation can find the passages that match a question by meaning rather than by exact words. Vector stores belong to a workspace. They are under Agentic ▸ Vector stores in the Orchestrator, and under Vector stores in the Agents menu. Vector stores need the agentic feature and the pgvector extension in the Orchestrator database. See Set up Agents. Create a vector store# You need vector-store.create in the workspace. Open Vector stores and select New vector store. Enter a Name, for example ap-policies. Agents and automations address the store by this name. Choose the Embedding model: text-embedding-3-small, text-embedding-3-large or text-embedding-ada-002. Keep the Dimension the Orchestrator expects: 1536 unless your administrator changed Ai:EmbeddingDimensions. A store with another dimension cannot hold documents. Choose the Distance metric: Cosine, Inner product or L2 (Euclidean). Use Cosine unless you know you need another. Select Create vector store. The store opens with three tabs: Documents, Search and Settings. Where embeddings are computedThe Orchestrator computes embeddings itself. A store created in the web app uses the deployment's AI settings (configuration section Ai: Ai:Provider set to OpenAiCompatible, with Ai:BaseUrl, Ai:ApiKey and Ai:EmbeddingModel). A store created through the API without a model uses the workspace's embedding:default model. Ask your administrator which one applies before you load a large set of documents. Load documents# Documents are added by an automation, not in the web app. A running job sends them to the runtime API with its execution token: Add or replace documentsHTTPCopyPOST /api/v1/runtime/vector-stores/ap-policies/documents Authorization: Bearer <execution token of the running job> Content-Type: application/json { "documents": [ { "documentId": "ap-policy-001", "content": "Invoice approval thresholds: invoices up to EUR 5,000 are approved by the AP clerk...", "metadata": { "title": "Approval thresholds", "section": "2.1" } } ] } A document with an existing documentId is replaced, so loading the same set again does not create duplicates. DELETE /api/v1/runtime/vector-stores/{name}/documents/{documentId} removes one document. A job run by a Robot may write to any store of its workspace. A design-time run (from Studio or the Python CLI) needs its user to hold vector-store.edit in the workspace. A common pattern is a small process (for example PolicyIndexer) that reads your policy documents and loads them, run by a trigger whenever they change. See Triggers and schedules. Search a store# Use the Search tab to check what a question finds. Open the store and select Search (or Search documents on the Documents tab). Enter a Query, choose how many Results (5, 10, 20 or 50) and select Search. The results list each document's Score, Document ID, Content and Metadata, best match first. From a Python automation, search with the runtime client: runtime.client().search_vector_store("ap-policies", "Which invoices need CFO approval?", top_k=5). See The runtime client. Use a store in an agent# On the agent's Tools tab, select the store under Vector stores. The agent can then search it as a tool, which is Low risk and runs without approval unless you switch Approve each call on. See Add tools. Delete a store# You need vector-store.delete. On Vector stores, select ⋮ ▸ Delete on the store, or open it and select Settings ▸ Delete vector store. Confirm with Delete vector store. Deleting a store deletes its documents. Agents that use it fail the publish check tools_resolvable until you remove the tool.

Create a vector store

Load documents

Search a store

Use a store in an agent

Delete a store