lance - SKILL.md Agent Skill

name: lance description: Lance columnar data format and LanceDB expertise for ML/AI workloads, vector search, embeddings, multimodal datasets, and .lance files. Use when the user mentions Lance, LanceDB, pylance, vector indexes, ANN search, IVF_PQ, IVF_HNSW_FLAT, HNSW, full-text search, dataset versioning, or migrating ML data from Parquet.

Use this skill to design, debug, and optimize Lance and LanceDB systems for ML-native data, embeddings, vector retrieval, and multimodal storage.

Lance SDK / file-format project: v7.0.0 stable, released May 27, 2026.
LanceDB Python: v0.33.0 stable, released May 28, 2026. v0.33.1-beta.0 exists; do not treat beta as the default stable recommendation.
Python packages: install pylance for the Lance format and lancedb for the embedded/vector database API.
Python support: LanceDB wheels target CPython 3.9+; confirm package metadata before pinning in production.
Recent additions: MemWAL/LSM write paths, materialized view API, formal catalog/namespace/table/index specs, branch/tag metadata maps, segmented BTree indices, distributed bitmap index build, FTS segment merging, serializable scalar-index caches, and nested blob export fixes.
LanceDB recent additions: IVF_HNSW_FLAT vector index type, model-backed FTS tokenizers, and sync create_index API alignment with async.

Identify whether the user is asking about Lance format (pylance, .lance, dataset versioning, storage layout) or LanceDB (lancedb, tables, search, indexes, reranking).
Prefer current stable versions unless the user explicitly asks about beta features.
For API examples, keep snippets small and verify names against installed docs/source when working inside a project.

Batch writes; avoid single-row insert loops unless latency is more important than fragment count.
Build vector indexes after enough data exists; small datasets often do not need ANN indexes.
Use scalar indexes for frequent filters and combine filters with vector search to reduce candidate sets.
Use full-text or hybrid search when natural language relevance matters, not only nearest-neighbor distance.
Use list_tables() rather than deprecated table_names().
Use session-level cache configuration rather than deprecated per-table index_cache_size.
On object storage, account for serial request chains; latency is often bounded by metadata, index, and data-page round trips.

Confirm latest lance-format/lance release before changing SDK guidance.
Confirm latest stable lancedb release on PyPI before changing Python guidance.