Skip to Content
We are live but in Staging 🎉
VectorAPI ReferenceOverview

Vector — API Reference

Package: dodil.k3.vector.v1 · Service: VectorService

13 RPCs across five functional groups. The single most important RPC is Search — it has three query shapes, three search modes, multi-collection support, optional Jina reranking, pre-filtering, and Milvus-native fast-lane tuning. That’s its own page below.

Sections

PageRPCsWhat it covers
EngineConfigureEngine · GetEngine · DeleteEngine · ListVBaseInstancesPer-bucket VBase backend lifecycle. Three modes — auto (K3 provisions), external (your own VBase), pick (existing service).
CollectionsAddVectorPipeline · AddVectorCollection · ListCollections · GetCollection · DeleteCollectionBoth creation modes — pipeline-mode (template-driven, schema lazy) and manual-mode (caller-managed schema, EXTERNAL embedding source).
SearchSearchThe deep dive — text / pre-embedded vector / file query shapes; VECTOR / HYBRID / AUTO modes; multi-collection compatibility; Jina rerank; metadata pre-filter; Milvus tuning.
VectorsInsertVectors · UpsertVectors · DeleteVectorsDirect writes to EXTERNAL-mode collections. Five dense element types (float / binary / float16 / bfloat16 / int8), optional sparse, optional text-for-BM25.
TemplatesListTemplatesVector-pillar template catalog — *_embedding_index (and matching _search) for text, code, visual, face, object.

Wire conventions

HTTP request bodyHTTP query / pathgRPC
Field namescamelCase (pbjson)snake_case (e.g. top_k, min_score)as in .proto
Enumswire-name strings ("SEARCH_MODE_AUTO")enum
int64 / int32JSON strings / numbers per pbjson rulesstrings / numberstyped
oneof request fieldsexactly one variant settyped
All fieldsalways emitted (no defaults dropped)

Endpoint roots

EndpointUse
https://k3.dev.dodil.io (staging) · https://k3.dodil.io (prod)HTTP
k3-grpc.dev.dodil.io:443 (staging) · k3-grpc.dodil.io:443 (prod)gRPC

Auth: bearer JWT in Authorization. See Conventions.

VBase touchpoints

K3’s vector engine is backed by VBase  (managed Milvus). For workflows K3 doesn’t surface directly:

SurfaceWhat it gets you
ConfigureEngine with mode: "external"Point the bucket’s vector engine at your own VBase cluster. Direct VBase API calls now hit the same backing store.
ListVBaseInstances (GET /admin/vbase-instances)Enumerate VBase services available in the org — feeds the pick mode service_id field.
Direct VBase APIFor raw Milvus features (custom indexes, partition lifecycle, advanced index params) K3 doesn’t expose. See VBase docs .

See also

  • Quickstart — end-to-end pipeline-collection RAG flow
  • Core Concepts — every type signature, enums, sparse modes, embedding types
  • CLI Guidedodil k3 vector store / collection / templates + dodil k3 search
  • Recipes — pipeline-collection, external-collection, multi-collection search, hybrid + rerank, multimodal