November 2026 Summaries

3 posts from LanceDB

Filter
Month: Year:
Post Summaries Back to Blog
Table formats address the mismatch between mostly immutable data files and datasets that need to change through appends, updates, and deletions, preventing teams from having to recreate versioning, indexing, and coordination through ad hoc scripts. They represent changes by writing new data, deletion, or index files and use metadata—called a manifest, snapshot, or log in different systems—to identify active files and explain their roles, enabling features such as access to older snapshots. A separate mechanism records commits and coordinates concurrent writers, either through filesystem operations that require little extra infrastructure or through a centralized service that can support different storage systems and transaction rates. Formats such as Iceberg, Hudi, Delta Lake, Lance, and DuckLake make different design choices, reflecting varied workloads and trade-offs rather than a single universally best approach. The article introduces these shared concepts as the starting point for a series on table-format features and implementation, while noting the author’s work at LanceDB may shape the terminology and examples.
Nov 16, 2026 2,350 words in the original blog post.
LanceDB’s update highlights LeRobot v0.7.0 support for training directly on nearly 300 TB of Lance-formatted robot data without downloads or video re-encoding, alongside an Apache DataFusion ASOF JOIN contribution for aligning camera frames with robot states. It also presents a LanceDB-based LLM pretraining workflow that retains raw text, deduplication flags, and token data in one table while performing sequence packing at read time. Releases across Lance, LanceDB, and lance-spark add full-text and MinHash indexing, materialized views, expanded blob handling, security safeguards for embedding configurations, distributed index building, improved query and I/O efficiency, versioned reads, and observability features. The update recognizes community contributions, recaps discussions on Blob v2, documentation, partial vector indexes, and index maintenance, and announces November and December 2026 events focused on robotics, physical AI, and cloud infrastructure.
Nov 12, 2026 1,483 words in the original blog post.
Augmetec uses LanceDB Enterprise as the data and retrieval foundation for LEIAA, its AI-powered platform for internal and regulatory investigations involving millions of documents. The system combines semantic vector search, full-text search, and metadata pre-filtering to help investigators find conceptually related evidence while retaining precise criteria and links to original documents. By indexing new evidence continuously and separating query-serving capacity from background indexing, the platform supports interactive investigation at scales exceeding millions of documents and hundreds of millions of vectors. Augmetec reports that LEIAA can surface high-risk material within minutes, return initial results in under five seconds, reduce review hours by more than 70%, and improve consistency by allowing related evidence to be assessed in context rather than divided among isolated reviewers.
Nov 10, 2026 1,369 words in the original blog post.