Search dense + sparse

Search dense and sparse vectors together, then combine the rankings with Reciprocal Rank Fusion.

POST /v1/tenants/{tenant}/vector/{table}/topk_hybrid

Combine dense and sparse rankingsMatching IDs connect both representations to the same source item.

Prepare both query representationsSupply dense coordinates and sparse positions with values.
Combine ranked candidatesUse Reciprocal Rank Fusion across the two searches.
Read the fused resultInterpret its score as fusion, not cosine similarity.

Before you start

Use your database bearer token.

Connection setup

Index dense and sparse vectors under the same record IDs before combining their search results.

Path parameters

Field	Type	Required
`tenant`	`string`	Required
`table`	`string`	Required

Headers

Field	Type	Required	Description
`x-oc-db-user`	`string`	Optional	Acting database user (role, grant and row-policy checks). Send with x-oc-db-key.
`x-oc-db-key`	`string`	Optional	API key of the x-oc-db-user user.

Request body

application/json

Limits and accepted values

rrf_k defaults to 60;
candidates (per-leg overfetch) defaults to max(4k, 40) and is never below k.
dense_mode: fast | high_recall.
`filter` is a strict-equality metadata map ({field: value});
`filter_expr` is the typed operator filter (e.g. {"tag": {"$eq": "a"}}) and cannot be combined with a non-empty `filter`;
`require_complete: true` turns an underfilled filtered answer into a 422.

Field	Type	Required
`candidates`	`integer`	Optional
`dense_dim`	`integer`	Required
`dense_metric`	`string`	Optional
`dense_mode`	`string`	Optional
`dense_query`	`number[]`	Required

Fields in dense_query

Field	Type	Required
`dense_query`	`number`	Optional

Field	Type	Required
`filter`	`object`	Optional
`filter_expr`	`object`	Optional
`k`	`integer`	Required
`require_complete`	`boolean`	Optional
`rrf_k`	`integer`	Optional
`sparse_dim`	`integer`	Required
`sparse_query_indices`	`integer[]`	Required

Fields in sparse_query_indices

Field	Type	Required
`sparse_query_indices`	`integer`	Optional

Field	Type	Required
`sparse_query_values`	`number[]`	Required

Fields in sparse_query_values

Field	Type	Required
`sparse_query_values`	`number`	Optional

Read the result

The returned score comes from Reciprocal Rank Fusion of dense and sparse rankings. It is not a cosine score. Both representations need matching record IDs to describe the same source item.

Response statuses

HTTP status reference

Status	What to do
200	OK
400	Correct the request before retrying.
401	Check the credential and Authorization header.
402	Check feature access or available credits.
403	Check tenant access, roles and permissions.
408	Reduce query work or adjust the timeout.
409	Resolve the conflict before retrying.
413	Split or reduce the request body.
415	Send the required Content-Type.
422	Correct the values or unsupported combination.
429	Honor Retry-After when present; back off.
499	The client cancelled the request.
500	Check the outcome before retrying a write.
501	Check deployment support for this operation.
503	Honor Retry-After when present; check write outcomes.

Request example

curl --include --request POST "${OC_BASE_URL}/v1/tenants/${OC_TENANT}/vector/demo_vectors/topk_hybrid" \
  --header "Authorization: Bearer ${OC_TOKEN}" \
  --header "Content-Type: application/json" \
  --data-binary '{
  "dense_dim": 3,
  "dense_query": [
    0.1,
    0.2,
    0.3
  ],
  "dense_metric": "cosine",
  "sparse_dim": 3,
  "sparse_query_indices": [
    0,
    2
  ],
  "sparse_query_values": [
    0.4,
    0.8
  ],
  "k": 5,
  "rrf_k": 60
}'

Response example

Examples use synthetic values to show the response format.

Response status

[
  {
    "id": "v1",
    "score": 0.0327868852
  },
  {
    "id": "v2",
    "score": 0.0161290323
  }
]

Illustrative RRF ranking with v1 in both legs; exact ranks and scores depend on the indexed corpus.

Download the OpenAPI specification for the complete response schemas.

Before you startCopy Anchor Link

Path parametersCopy Anchor Link

HeadersCopy Anchor Link

Request bodyCopy Anchor Link