qdrant changelog

0.3.1

Released 2026-10-21.

Changes

Give qdrant memory tools time to start (#23)

Loading the embedding model takes about 90 seconds on one CPU even from the cache, so the liveness probe killed the server before it listened. A startup probe now allows ten minutes, which also covers the first model download.

0.3.0

Released 2026-10-21.

Changes

Let qdrant memory pick its model (#22)

qdrant 0.3.0 runs the image of upstream's own Dockerfile, which downloads the embedding model at first use: the model now goes to a volume once, so later starts need no internet, and a random UID gets writable cache and home directories there. mcp.embeddingModel selects the FastEmbed model, for example the multilingual MiniLM for teams that write in more than one language.

0.2.0

Released 2026-10-21.

Changes

Add memory tools to chart qdrant (#21)

qdrant 0.2.0 runs the official mcp-server-qdrant next to Qdrant when mcp.image is set: qdrant-find and qdrant-store on streamable HTTP, on a Service an MCP gateway discovers by its appProtocol. Installing the chart gives a project's agents a shared long-term memory and removing it takes the tools away, with nothing to change on the agents.

The tool descriptions carry the usage rules (search at the start of a task, save decisions, preferences and lessons), because that is the only guidance that exists exactly when the tools do. The server can reach its own Qdrant and DNS only.

0.1.0

Released 2026-10-21.

Changes

Add chart qdrant (#20)

One Qdrant vector database per release namespace, for RAG and agent memory. Based on the chart Amit Choubey wrote for chat.grn.cloud, moved here from openshift-config so tenants install it from the marketplace.

The API key is generated once and kept on upgrade, or read from a secret store; the health endpoints need no key.