Hybrid vector search in your bucket.

Ask your bucket a question. Get an answer ranked by meaning, not just by keyword — hybrid dense + sparse retrieval with RRF reranking, multimodal across text, image, audio, and video, attached to the K3 Objects prefix you already use.

Your bucket is already indexed.

Attach a vector pipeline to a K3 Objects prefix once. Every PUT after that gets chunked, embedded, and upserted into the collection automatically — no embedder service to host, no glue code to write.

Why hybrid wins.

Dense vectors catch meaning. BM25 catches exact tokens. Reciprocal-rank fusion combines both rankings into one — for the same query, hybrid consistently surfaces the right document and the right paragraph.

Grounded answers, in three hops.

Retrieve from your collection, hand the context to an Ignite Models endpoint, return the answer with citations that link back to the source objects in your bucket.

Built for four jobs.

RAG, semantic search, deduplication and clustering, agent long-term memory — all the same collection type, all bucket-attached.

Speaks Milvus. Plus a few REST calls.

Use pymilvus or the Milvus REST API unchanged — your existing search code works against a Dodil endpoint. The dodil CLI adds collection management and the pipeline DSL on top.

# your existing pymilvus code works unchanged
from pymilvus import MilvusClient
 
client = MilvusClient(
uri="https://vec.uk.dodil.io",
token=DODIL_TOKEN,
)
 
results = client.search(
collection_name="contracts",
data=[query_embedding],
limit=4,
search_params={"mode": "hybrid"},
output_fields=["source", "page"],
)

Pay per query. Not per cluster.

No idle cluster. No per-collection fee. You pay only for the vector reads, writes, and the bytes you actually keep.

Three primitives. One bucket.