dodil data vsearch

Vector KNN search over a table’s vector column, executed on the tables data plane through the typed TablesVector.QueryVectors gRPC facet — not over the Qdrant wire. Alias: dodil data knn.

A “collection” in tabled is a table with a vector column, so this command wants a table and a column, not a collection id:

dodil data vsearch -b BUCKET -t TABLE --column COLUMN \
  (--vector "0.12,0.03,…" | --text "some query") [--metric M] [--top-k N]

Exactly one of --vector or --text is required; with neither, the command errors. --metric is mandatory on the wire — the plane rejects METRIC_UNSPECIFIED — so the CLI always sends one, defaulting to cosine.

--text needs the unified dodil CLI. It embeds through Ignite Models via an injected hook; the standalone datak3 binary leaves that hook unset and --text errors with “embed yourself and pass --vector”.

Search

# By text — one Ignite embed call, then the KNN
dodil data vsearch -b kb-prod -t chunks --column embedding \
  --text "refund policy" --top-k 10
 
# By vector — you already have the floats
dodil data vsearch -b kb-prod -t chunks --column embedding \
  --vector "0.12,0.03,…" --metric cosine --top-k 10

from qdrant_client import QdrantClient
 
qd = QdrantClient(url="https://qdrant.uk-lon-1.dodil.io", api_key="kb-prod")
 
hits = qd.search(
    collection_name="chunks",       # the table
    query_vector=[0.12, 0.03, ...], # the `vector` column
    limit=10,
    with_payload=True,
)
for h in hits:
    print(h.id, h.score)

-- pgvector operators over the Postgres wire: <=> cosine, <-> L2, <#> inner product
SELECT id
FROM chunks
ORDER BY embedding <=> '[0.12, 0.03, ...]'
LIMIT 10;

Output

Table output is id · score rows plus a hit count. -o json returns the CLI’s own envelope — not a Qdrant-wire response, and no payloads:

{
  "matches": [
    {"id": "chunk-41", "score": 0.1832},
    {"id": "chunk-07", "score": 0.2114}
  ],
  "count": 2
}

With --include-values, each match also carries "values": [...].

score is a distance, not a similarity — lower is closer. VectorMatch.score on dodil.tables.v1 is the raw metric distance (dodil-tables/proto/api/tables.proto:1104-1112). Don’t compare it against the RRF fusion scores that POST /:bucket/search/vector returns.

For payloads and metadata filters, use a stock Qdrant or Pinecone client — see Connect & wire adapters. The console’s Connect panel generates ready-to-paste snippets per language.

Text search with server-side embedding

vsearch --text embeds client-side — pure KNN semantics, no hybrid, no rerank. For server-side embedding over pipeline-mode collections (and the cross-group RRF merge), use the HTTP route instead:

curl -sS -X POST "https://api.data.dodil.io/kb-prod/search/vector" \
  -H "Authorization: Bearer $DODIL_TOKEN" \
  -H "Content-Type: application/json" \
  -d '{"text": "refund policy", "collectionNames": ["docs"], "topK": 10}'

Collection lifecycle

Creating, listing, and deleting collections stays on the control plane — and all three key off the pipeline id:

dodil data vector collection list   -b kb-prod
dodil data vector collection add    my-texts -b kb-prod -t text_embedding_index
dodil data vector collection delete <pipeline-id> -b kb-prod

See also

• Vector → Core Concepts — the SQL vector data model, metrics, and dispatch
• API Reference → Search — the managed POST /:bucket/search/vector route
• Connect & wire adapters — Qdrant / Pinecone / Postgres endpoints + credentials
• Vector → Wire Compatibility — the Qdrant / Pinecone adapter surface