Qdrant
A RocketRide vector store node backed by Qdrant, for storing embedded documents and retrieving them by semantic or keyword search.
About Qdrant
Qdrant is an open-source vector database built for similarity search at scale. It stores high-dimensional embeddings and returns the closest matches to a query vector, with filtering and payload support, and runs self-hosted or as a managed service (Qdrant Cloud).
What it does
Stores pre-embedded document chunks in a Qdrant collection and retrieves them by semantic (vector) similarity or full-text keyword search. Works with both self-hosted Qdrant and Qdrant Cloud. The node is also a tool node (classType: ["store", "tool"] with the invoke capability), so agents in the same pipeline can use the eight functions documented under As a tool, including document retrieval, statistics, trash and restore, subject to the tool-exposure and write-permission settings.
Uses the official qdrant_client Python SDK over REST (prefer_grpc=False, 60-second timeout). If the configured host has no scheme, http:// is prepended automatically.
Documents must be run through an embedding node before reaching this node: the store does not compute embeddings on the data lane and raises an error if a chunk arrives without one.
The collection is created automatically on first write, including payload indexes on meta.nodeId, meta.objectId, meta.parent, meta.permissionId, meta.isDeleted, meta.isTable, and a full-text index on content (word tokenizer, lowercase, token length 2-15). Writing chunk 0 of an object first deletes all existing points with the same objectId, so re-ingesting a document replaces it rather than duplicating it. Upserts are batched: a flush happens every 500 points or when the accumulated payload exceeds the payload limit (32 MiB by default).
Example pipelines
End-to-end RAG pipeline
webhook → parse → preprocessor_langchain → embedding_transformer → qdrant
chat → embedding_transformer → qdrant → prompt → llm_openai → response_answers
Documents are parsed, chunked, embedded, and stored in Qdrant. Chat questions use the same embedding and Qdrant retrieval path before the prompt and LLM produce an answer.
Question answering over stored documents
webhook → embedding_transformer → store_qdrant → llm_anthropic → response
The question is embedded, this node enriches it with the most relevant stored
documents (questions → questions lane), and the LLM answers from that
retrieved context.
Agent with searchable memory
An agent with this node connected as a tool: it calls search for context
when it decides it needs it, and upserts new facts as it works — a knowledge
base the agent reads and writes on demand instead of retrieval running on
every question.
Lanes
| Lane in | Lane out | Description |
|---|---|---|
documents | — | Ingest pre-embedded documents into the collection |
questions | documents | Return matching documents |
questions | answers | Return matching documents as an answer |
questions | questions | Enrich the question with matching documents for downstream nodes |
As a tool
Pipeline agents address the functions as <node id>.<function>. The configured serverName (default qdrant) labels tool-response messages. The available functions are search, objdir, get, stats, upsert, delete, soft_delete, and restore; write functions require allow_write, and all functions require expose_tools.
search
Semantic similarity search over the collection.
| Argument | Type | Description |
|---|---|---|
query | string | Required. The search query text. |
top_k | integer | Maximum results to return (default 10, capped at 100). |
filter | object | Optional metadata filter; supports objectId, nodeId, and parent keys. |
Requires an embedding provider bound to the node (the embedding sub-block in the configuration). If semantic search fails, the tool falls back to keyword search. Returns {"results": [{content, score, metadata}], "total": n} on success or {"success": false, "error": "..."} on failure.
objdir
List the distinct parent paths of active documents in the collection as a map of parent path to one object ID each. Takes optional parent (a text match on the stored parent, not an exact path), offset and limit (default and cap 1000). offset and limit count documents, not paths: documents that share a parent path appear as one entry, so a page can hold fewer than limit entries while more follow. Qdrant's scroll offset is a point-ID cursor, so each page reads offset + limit points from the start of the collection, in requests of at most 1,000 points, and skips the first offset.
get
Return the chunks of one document. Takes a required object_id and optional offset and limit (default and cap 100); chunks come ordered by chunk ID with total, has_more and truncated. An object ID the collection does not hold, or one in the trash, returns {"success": false, "error": "..."}.
stats
Return the collection's point count as Qdrant reports it, document_count (null on this store, which does not list documents), the embedding model name and the vector size.
upsert
Add or update documents in the collection.
Each item in the documents array requires content and object_id (used for deduplication) and may include metadata (keys nodeId/parent/chunkId are stored; nodeId defaults to "vectordb_tool", parent to "/", and chunkId to the position among the chunks sent for that object ID in the call, 0 for the first; an explicit chunkId may not repeat within the call), plus optional pre-computed embedding and embedding_model to skip automatic embedding. Without a pre-computed vector the node's bound embedding provider is used; if none is configured the call fails. Returns {"success": true, "count": n, "skipped": n}.
delete
Delete documents by ID.
Takes a required object_ids string array and removes all points whose meta.objectId matches. Returns {"success": true, "deleted_count": n}.
soft_delete and restore
Move documents to the trash and back by object ID. Both take a required object_ids string array; soft_delete marks every point whose meta.objectId matches as deleted, so it leaves search and the read tools until restore clears the mark. Both return {"success": true, "object_ids": [...], "matched_chunks": null}, since this store does not count the points it marks. A document removed with delete cannot be restored. Both need a Qdrant server of 1.8 or newer, since they update only the nested meta.isDeleted field.
Profiles
Default: Qdrant cloud server (cloud).
| Profile | Default host | Default port |
|---|---|---|
| Qdrant cloud server (default) | (your Qdrant Cloud endpoint) | 6333 |
| Your own Qdrant server | localhost | 6333 |
Configuration
Choose the host type (cloud or self-hosted), point the node at your Qdrant endpoint, and name the collection. Most deployments only ever touch the host, API key, and collection fields.
Tool Server Name
The name defaults to qdrant and labels this store in tool-response messages.
Pipeline agents address all eight functions documented above by node ID, for
example <node id>.search or <node id>.restore; changing this label does not
change those addresses.
Authentication
Self-hosted Qdrant typically needs no credentials: leave apikey empty. For Qdrant Cloud, set host to your instance endpoint (<instance>.<region>.qdrant.io) and provide the cluster API key in apikey. The key is passed as the api_key argument of the Qdrant client.
Notes
Scoring and relevance
Raw Qdrant scores are rescaled to [0, 1] before filtering. For Cosine similarity the rescaling is (score + 1) / 2; for other metrics a sigmoid is applied. Results below a hard floor of 0.20 are always dropped as noise, and the configured retrieval score threshold is applied on top of that in the rescaled space (applied by the base store, not passed to Qdrant as score_threshold).
The similarity metric defaults to Cosine; Euclid, Dot, and Manhattan are also accepted. Any other value raises a configuration error at startup. Semantic search uses exact search (SearchParams(exact=True)) and does not support a non-zero result offset; keyword search and plain filtered gets use Qdrant scroll and do support offsets.
Deletion of source files is handled as a soft delete: chunks are marked meta.isDeleted: true and excluded from all searches by default. They are only returned when the filter explicitly asks for deleted documents. A document that reappears is marked active again.
Save-time validation
When the configuration is saved, the node validates it with a fast read-only probe (get_collections, 10-second timeout):
- The collection name must match
^[A-Za-z0-9._-]{1,255}$: no/or spaces. - Port
0is rejected explicitly. - Scheme is inferred when the host has none:
httpforlocalhostor127.*,httpsfor*.qdrant.iohosts or port443. - Qdrant Cloud with a wrong port does not return an HTTP status: the probe simply times out after 10 seconds and surfaces a timeout warning.
Language support and rehydration
- Qdrant's full-text tokenizer does not support all languages out of the box. Chinese, Japanese, and Korean are not enabled in the default Qdrant build and require building Qdrant from source with the
--features multiling-chinese,multiling-japanese,multiling-koreanflags. This only affects keyword search over thecontentpayload; semantic search uses the embeddings you provide and is unaffected. - Rendering a full document rehydrates all of its chunks in
chunkIdorder, fetching them in batches of up to 32 MiB and joining them, so out-of-order storage is handled transparently.
Upstream docs
Schema
| Field | Type | Description | Default |
|---|---|---|---|
qdrant.profile | string | Type of Qdrant host Connect to... | "cloud" |
qdrant.provider | string | const: "qdrant" | |
qdrant.serverName | string | Tool Server Name Label for this store in tool-response messages (for example, a refused write names qdrant.delete). Pipeline agents address the available functions as node-id.function; apps use the function name and node ID. Changing this label does not change those addresses. See the node README for the available functions and their tool-exposure and write-permission settings. | "qdrant" |
vector.cloud.host | Enter the server IP address e.g. | ||
vector.cloud.port | 6333 | ||
vector.local.host | "localhost" | ||
vector.local.port | 6333 |
Dependencies
grpciogrpcio-toolsportalockerpydanticurllib3httpxqdrant_client==1.17.1numpy
