Guide

Vector Store Endpoints

Eden exposes shared endpoint configuration for vector stores used by migration and retrieval workflows. The grouped vector-store endpoint page covers:

Qdrant
Milvus / Zilliz
Elasticsearch
OpenSearch
ChromaDB
LanceDB
Turbopuffer
Apache Solr
Mixpeek Vector Store

These endpoint configs use the split credential shape:

json

{
  "target": {
    "url": "https://vector-store.example.com",
    "collection": "documents"
  },
  "read_credentials": {
    "api_key": "..."
  }
}

Use collection, index, or namespace depending on the provider. Optional database is available for providers such as ChromaDB and Milvus / Zilliz.

Migration credential roles

Historical migration adapters use credentials by role rather than treating one API key as universal:

read_credentials are required on a historical source and for target-side

validation.

write_credentials are required on a historical target.
admin_credentials are required only when Eden must create a missing

collection or index.

Credentials, configured headers, and credential-bearing URL parameters are not copied into migration batches, checkpoints, or semantic write events.

Executable preview paths

Endpoint configuration support is broader than migration support. The current historical previews are deliberately narrow:

Qdrant to Qdrant preserves collection/vector configuration, copies points by

server scroll cursor, performs acknowledged idempotent upserts, and validates counts, dimensions, and sampled content. The migration must use replace conflict semantics. A missing target collection requires admin_credentials. Custom-sharded collections are unsupported. Collections with payload indexes require a pre-created target with matching collection configuration and payload schema. Collection metadata, named/unnamed vector shape, and payload-only points are preserved.

OpenSearch and Elasticsearch support same-kind copies and both cross-provider

directions through the search/document family. Cross-provider migrations require explicit identity index routes. Hidden and system indices are excluded unless the endpoint explicitly names one. Custom analysis chains, disabled _source, provider lifecycle/plugin settings, vector mapping conversions, target alias collisions, index renames, and field conversions are planning blockers. Restart recovery performs a safe full idempotent rescan; stable cursor resume is not yet advertised.

Both preview paths are offline-only. Stop source writes before starting, pass source_quiescence_confirmed: true to /migrate and every /resume, keep the source quiesced, and pass source_quiescence_maintained: true to /complete. These values are operator attestations; Eden does not yet hold a durable provider write fence.

Pinecone, Weaviate, Milvus/Zilliz, ChromaDB, LanceDB, Turbopuffer, Solr, and Mixpeek Vector Store remain endpoint configurations only for this migration surface. Their presence in the endpoint catalog does not authorize historical copy, reverse copy, capture, replay, or cutover.

HTTP live-write migration and gRPC are not published capabilities. Collection administration, scripted search updates, update-by-query, and other unsupported mutations must be handled outside the preview rather than assumed replayable.