Skip to main content

Qdrant

View as Markdown

Qdrant

Qdrant is a vector database for storing and querying embeddings. In RocketRide it acts as both the retrieval backbone for RAG pipelines and as an agent tool for semantic search, upserts, and deletions.

When to use Qdrant​

  • RAG — embed documents, store vectors, retrieve relevant chunks at query time.
  • Semantic search — find the most relevant items across a large corpus based on meaning rather than keywords.
  • Hybrid search — Qdrant supports combining vector similarity with full-text keyword scoring in a single query.
  • Agent memory — an agent can call the Qdrant tools directly to search or update the store as part of a reasoning loop.

Configuration​

Profiles​

The qdrant node ships with two built-in profiles:

ProfileUse when
cloud (default)Qdrant Cloud (*.qdrant.io)
self-hostedLocal Docker or on-premises
{
  "id": "store_1",
  "provider": "qdrant",
  "config": {
    "profile": "self-hosted",
    "serverName": "localhost",
    "collection": "my-docs"
  }
}

For Qdrant Cloud, set serverName to your cluster hostname and provide an API key via the apikey field.

Collection management​

The node creates the collection automatically if it does not exist. Vector dimensions are inferred from the first batch of embeddings — make sure the embedding model you use in the pipeline is consistent across all runs against the same collection. Mixing text-embedding-3-small and text-embedding-3-large in the same collection will fail because the dimensions differ.

Validation​

On pipeline startup, the node validates the collection name (alphanumeric and hyphens only), the port number, and connectivity. A failed validation prevents the pipeline from starting.

Wiring: document ingestion + retrieval​

The qdrant node has two input roles depending on the lane:

  • documents lane — receives embedded chunks from an embedding node and upserts them into the collection.
  • questions lane — receives a query, runs a similarity search, and emits the top-K results back as questions with context injected (for an LLM to consume) or as documents (for downstream nodes).

A single qdrant node can handle both at the same time:

{
  "id": "store_1",
  "provider": "qdrant",
  "config": { "profile": "self-hosted", "serverName": "localhost", "collection": "docs" },
  "input": [
    { "lane": "documents", "from": "embed_1" },
    { "lane": "questions", "from": "source_1" }
  ]
}

New documents are upserted as they arrive. Questions trigger a retrieval immediately, using whichever vectors are already in the collection.

Agent tools​

When connected to an agent via a control connection, the qdrant node exposes three tools. Tool names are namespaced by the node's serverName config (default qdrant), so they appear to the agent as qdrant.search etc. Change serverName when running multiple Qdrant nodes in one pipeline so their tool names do not collide.

ToolWhat it does
qdrant.searchSemantic similarity search over the collection. Returns the top-K chunks with scores and metadata.
qdrant.upsertInsert or update a document by ID. The agent provides the text; the node embeds and stores it.
qdrant.deleteRemove a document from the collection by ID.

The agent calls these tools in its reasoning loop — useful when the agent needs to both answer questions and update the knowledge base during the same session.

Full RAG pipeline example​

See the RAG Pipeline example for a complete, annotated .pipe file.