LanceDB as Default Vector Database Backend

[kheiss-uwzoo]

PR Summary: LanceDB as Default Vector Database Backend

Overview

Documentation updates establish LanceDB as the default vector database backend for NeMo Retriever extraction, with Milvus documented as a fully supported alternative. Content from the LanceDB product brief is integrated into the existing extraction docs in logical places.

Summary of Changes

1. Data Store Documentation (extraction/data-store.md)

• Overview — States LanceDB as default vector DB and Milvus as alternative; clarifies that vdb_upload targets the configured backend.
• Why LanceDB? — New section: Lance columnar format, IVF_HNSW_SQ index, embedded runtime, and why they improve latency.
• Upload to LanceDB (default) — New section with:
  • Python LanceDB operator example (uri, table_name, index_type, hybrid).
  • Test harness configuration (tools/harness/test_configs.yaml: vdb_backend, hybrid, sparse).
  • Environment variables: VDB_BACKEND, HYBRID.
• Hybrid search (LanceDB) — New section: BM25 FTS + vector, RRF, recall table (+0.5%–3.5%), latency note, one-time FTS index build.
• Infrastructure: LanceDB vs Milvus — New comparison table (runtime, external services, Docker profile, index type, hybrid approach, persistence).
• Upload to Milvus — Kept as alternative; clarified that you can continue using Milvus by setting vdb_backend: milvus or using the existing Milvus API.

2. Product-Level and Overview Docs

• extraction/overview.md — Vector DB sentence updated to “LanceDB by default, or Milvus”.
• index.md — “Embedding + Indexing” bullet updated to “LanceDB (default) or Milvus”.
• extraction/support-matrix.md — Retrieval bullet updated to “LanceDB (default) or Milvus”.

3. Python API and Environment

• extraction/nv-ingest-python-api.md — vdb_upload table and example caption updated to mention LanceDB as default and Milvus via milvus_uri.
• extraction/environment-config.md — New “Vector Database (Retrieval) Environment Variables” table: VDB_BACKEND, HYBRID.

4. Quickstarts

• extraction/quickstart-guide.md — Step 1f and profile table updated so default is LanceDB (no --profile retrieval); retrieval profile is for Milvus. Added tip that docker compose up without --profile retrieval uses LanceDB. Container/URI note updated for when Milvus is used.
• extraction/quickstart-library-mode.md — Install step: default is LanceDB (no milvus-lite required); optional Milvus path described. Ingestion example uses LanceDB by default (no milvus_uri). Step 3 retrieval example described as “when using Milvus”; LanceDB retrieval pointed to data-store.

5. FAQ and Release Notes

• extraction/faq.md — “Where does it ingest to?” updated for LanceDB default and Milvus supported.
• extraction/releasenotes-nv-ingest.md — LanceDB bullet updated to “LanceDB is now the default vector database backend” and linked to Data Upload.

Files Modified

File	Changes
docs/docs/extraction/data-store.md	LanceDB as default, “Why LanceDB?”, LanceDB upload + API + harness, hybrid search, infrastructure table, Milvus as alternative
docs/docs/extraction/overview.md	Vector DB sentence (LanceDB default, Milvus)
docs/docs/index.md	Embedding + indexing bullet (LanceDB default or Milvus)
docs/docs/extraction/support-matrix.md	Retrieval bullet (LanceDB default or Milvus)
docs/docs/extraction/nv-ingest-python-api.md	vdb_upload description and example (LanceDB default, Milvus option)
docs/docs/extraction/environment-config.md	VDB env vars: VDB_BACKEND, HYBRID
docs/docs/extraction/quickstart-guide.md	Default = LanceDB; retrieval profile for Milvus; container/URI note
docs/docs/extraction/quickstart-library-mode.md	Default = LanceDB; optional Milvus; ingestion/retrieval examples
docs/docs/extraction/faq.md	Ingest destination (LanceDB default, Milvus)
docs/docs/extraction/releasenotes-nv-ingest.md	LanceDB as default backend, link to Data Upload

Source

Content drawn from the LanceDB product brief (LanceDB Replaces Milvus as the Default Vector Database Backend), including rationale (format, index, embedded runtime), hybrid search (BM25 + vector, RRF) and recall table, test harness and env configuration, programmatic LanceDB usage, and infrastructure comparison (LanceDB vs Milvus).

Branch

LanceDB documentation branch.