Skip to content

Qdrant

Qdrant vector database with optional memory tools for AI agents (mcp-server-qdrant) in the release namespace, protected by an API key
Chart qdrant
Version 0.3.1
Application version v1.19.2
Kind Application
Source charts/qdrant

The values are listed under Values, the changes per version under Changelog.

Purpose

Qdrant is a vector database: it stores embeddings for retrieval-augmented generation (RAG), semantic search and agent memory. The chart runs one upstream Qdrant in the release namespace with a data volume, behind an API key. Clients in the same project, and in namespaces you allow, reach it at port 6333 (HTTP) and 6334 (gRPC); there is no public address.

With mcp.image set, the chart also runs the official mcp-server-qdrant next to it: long-term memory tools for AI agents (qdrant-find, qdrant-store) on streamable HTTP. Its Service carries appProtocol: agentgateway.dev/mcp and the annotations agentgateway.dev/mcp-path and agentgateway.dev/mcp-target-name, so an MCP gateway with a label selector discovers it and the project's agents get the tools as <targetName>_qdrant-find and <targetName>_qdrant-store. Uninstalling the release removes the tools again; agents need no change either way. The tool descriptions tell the model when to search and when to save.

Install

helm repo add grncloud https://apps.openshift.eu
helm install memory grncloud/qdrant \
  --set mcp.image.repository=<registry>/mcp-server-qdrant --set mcp.image.tag=<tag>
  • persistence.size (default: 5Gi): size of the data volume.
  • persistence.storageClass (default: empty): StorageClass of the data volume. Empty: the cluster default.
  • secrets.type (default: kubernetes): externalSecret reads the API key from a secret store.
  • secrets.apiKey (default: empty): API key. Empty: generated at the first install and kept on upgrade.
  • connection.configMap (default: true): ConfigMap <release>-qdrant-connection with QDRANT_URL, QDRANT_GRPC_URL and the name of the key Secret.
  • networkPolicy.ingressNamespaces (default: empty): label selectors of further namespaces that may connect to Qdrant.
  • mcp.enabled (default: true): run the memory tools. They render only with mcp.image.repository and mcp.image.tag set.
  • mcp.image (default: empty): an image of mcp-server-qdrant. Upstream publishes none; build the repository's own Dockerfile at a release tag.
  • mcp.collection (default: memory): collection that holds the memories.
  • mcp.embeddingModel (default: sentence-transformers/all-MiniLM-L6-v2): FastEmbed model, CPU only. A collection keeps the model it was created with; for mixed languages use sentence-transformers/paraphrase-multilingual-MiniLM-L12-v2 (384 dimensions, about 750Mi of memory in use).
  • mcp.modelCache.enabled (default: true), mcp.modelCache.size (default: 1Gi): volume the model downloads to once; later starts need no internet.
  • mcp.targetName (default: memory): tool prefix on the gateway.
  • mcp.findDescription, mcp.storeDescription: tool descriptions; the defaults say to search at the start of every task and to save decisions, preferences and lessons learned.
  • mcp.ingressNamespaces (default: empty): label selectors of the namespaces that may call the tools, for example the gateway's.
  • mcp.allowInternet (default: true): let the tool server reach the internet to download its embedding model from Hugging Face; false once the model is in the cache or built into the image.

Clients of Qdrant itself send the key from Secret <release>-qdrant-api-key (key api-key) in the api-key header:

curl -H "api-key: $QDRANT_API_KEY" "$QDRANT_URL/collections"

A tool that renders the chart without access to the cluster, such as Argo CD or helm template, generates a new key on every render: set secrets.apiKey there, or use secrets.type=externalSecret.

Delivery instructions

  • Install one release per project that needs its own vectors; data is not shared between releases, so each team's agents share one memory.
  • helm uninstall keeps the data volume (a StatefulSet volume claim). Delete the claim data-<release>-qdrant-0 to remove the data.
  • With secrets.type=externalSecret, create the vault item named in secrets.externalSecret.items.apiKey (default <release>-qdrant-api-key) with the key in the password field, and set secrets.externalSecret.storeName to your store.
  • The tool server has no login of its own: anyone who reaches its port can read and write the memory. Keep mcp.ingressNamespaces to the gateway, and let the gateway authenticate callers.