Qdrant

Workspace 0.4.0 inspects Qdrant collections and runs filtered vector queries through the REST API. Qdrant 1.19.2 with read-only and administrator API keys is the tested baseline. The Query API needs Qdrant 1.10 or newer; other versions and managed Qdrant Cloud deployments have not been verified.

Connect

Choose Qdrant, enter the host and REST port (6333 by default), and provide an API key if required. The key uses the system keychain, separately from saved connection settings. Use a read-only key for inspection. The gRPC port is not supported.

Set Collection when a key is restricted to one collection. Workspace then skips global collection discovery and refuses console commands for another collection. Collection scoping and unavailable cluster diagnostics are covered by REST contract tests; live managed-service JWT authentication has not been tested.

Choose Require for HTTPS and set the port used by that endpoint. Certificates and hostnames are verified with system roots or a private PEM CA bundle. HTTPS through an SSH tunnel requires the original database DNS hostname. The CA file is read on the computer running Workspace.

Views

The collection picker lists up to 200 names. Overview reads details for up to 20 collections and labels partial totals. Point and indexed-vector counts are server metadata and can lag behind writes; named and sparse vectors can also make those counts differ. The latency chart is a measured request, not CPU, memory or throughput. Workspace does not measure Qdrant recall.

Filtered vector search

This example uses the eight-dimensional vectors in the local merchant_profiles fintech demo. Supply vectors with the dimensions and meaning used by your own collection.

{
  "collection": "merchant_profiles",
  "operation": "query",
  "query": [0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8],
  "filter": {"must": [{"key": "country", "match": {"value": "NL"}}]},
  "limit": 10,
  "params": {"exact": true}
}

Query accepts a dense vector, a sparse { "indices": [1, 9], "values": [0.2, 0.4] } vector, or an existing numeric/UUID point ID. Set using for a named vector. Native Qdrant payload filters are passed to the server. Supported search parameters are exact and hnsw_ef (1–10,000).

Scores follow the collection's distance metric. They are not recall percentages or probabilities.

Browse, count and retrieve

{"collection":"merchant_profiles","operation":"scroll","limit":100}

If more points exist, the console shows the next scroll offset as JSON text. Copy that value into offset without converting it to a JavaScript number; large numeric IDs stay exact.

{"collection":"merchant_profiles","operation":"count","exact":true}

Count defaults to exact: false; request an exact count explicitly when needed. Retrieval uses operation: "retrieve" and an ids array. Numeric and UUID IDs are supported. For a collection-scoped connection, collection may be omitted.

Limits and exports

Only query, scroll, count and retrieve are available. Writes, collection management, arbitrary URLs, remote inference and prefetch/hybrid pipelines are not supported. Safe Mode cannot be disabled for this adapter.

Query and scroll default to 100 results and accept limits from 1 to 1,000. Retrieve accepts up to 1,000 IDs. Payloads are included by default; vectors are omitted unless with_vector is true. Input is capped at 64 KiB and responses at 8 MiB. Requests use an eight-second server timeout and a ten-second client timeout. Stop waiting ends the client wait and does not guarantee server cancellation.

CSV and JSON exports contain the displayed table. IDs and other cell values are strings, with payload/vector objects encoded as JSON text in their cells. This preserves large IDs but is not a collection backup or a portable vector migration file. See query history and export. For migration between embedded Recern Vector and Qdrant, use the separate Recern Vector portable API.