LanceDB

3/3 Talks from Lei Xu (CTO @ LanceDB), Goutam Venkatramanan (Software Engineer @ @anyscalecompute), and Hubert Yuan (Software Engineer @ @ExaAILabs) If you’re building AI search or large-scale retrieval systems, this is a concrete look at how it runs in production. Register: events.builder.aws.com/d/73q3y1

@anyscalecompute @ExaAILabs 2/3 How the system is built: - Lance for storing embeddings + raw data with fast random access - Ray for distributed ingestion and embedding pipelines - Continuous updates to search signals without rebuilding the index

1/3 What does it take to run search over billions of documents *and* keep it continuously updated? Join LanceDB, @anyscalecompute, and @ExaAILabs on March 31 in SF for a deep dive into the infrastructure behind Exa’s web-scale AI search engine.

3/3 Result: less back-and-forth from visual reference to implementation. Shoutout to the @ByteDanceOSS team for the writeup: lancedb.com/blog/openclaw-…

1/3 LanceDB, @openclaw, and Seed 2.0 let you go from a few screenshots to a working frontend without writing specs. The bottleneck isn’t generation—it’s retaining design intent across iterations.

2/3 Screenshots capture appearance, but not the reasoning behind decisions. This system treats that workflow as data: - OpenClaw captures and updates reference designs - LanceDB stores screenshots, notes, code, and revisions - Seed 2.0 generates code directly from that context

@changhiskhan @devstein64 @dosu_ai @joaomdmoura @crewAIInc @harjotsgill @coderabbitai Agentic AI is moving from demos to real engineering systems. We’ll discuss the infrastructure and architectural tradeoffs behind real-world agent systems. Hosted by Lancedb & @dosu_ai. March 26, SF. Register: luma.com/ga3zi7sx

Join us March 26 in SF for a panel with leaders building the next generation of AI-native developer tools: @changhiskhan — CEO, @lancedb @devstein64 — CEO, @dosu_ai @joaomdmoura — CEO, @crewAIInc @harjotsgill — CEO, @coderabbitai

In other words: unstructured data → embeddings → searchable knowledge That’s exactly the workflow many teams are building on LanceDB today — storing and querying multimodal data without stitching together multiple systems. Super awesome to see Lance & LanceDB included here, and shoutout to our partners on the @nvidia cuVS team!

~90% of the world’s data is unstructured — PDFs, videos, images, audio, documents. “Just as AI was able to solve multimodality problems — perception and understanding — you can use that same technology… to read a PDF, understand its meaning, and embed it into a structure we can search and query.”

Did you spot Lance & LanceDB in Jensen Huang's keynote at @nvidia GTC today? 👀 Jensen highlighted one of the biggest shifts in AI infrastructure: unlocking unstructured data.

@openclaw 3/3 Full walkthrough of OpenClaw + LanceDB memory: lancedb.com/blog/openclaw-…

2/3 Why this pairing works well architecturally: - Runs as an embedded library, not a separate database service - Embeddings, metadata, and indexes stored together in the same table - Vector + full-text + structured filtering over the same memory dataset - Append-friendly storage that matches how agents write memories

1/3 @OpenClaw agents persist memory across sessions. Preferences, facts, and prior decisions can be written during one interaction and retrieved the next time the agent runs. In the #OpenClaw ecosystem, @lancedb is quickly becoming the default memory layer.

Taken from last year's team offsite! So it only makes sense that LanceDB is the default memory layer for @openclaw 🦞

@openclaw @tonbistudio Vector + BM25 + Jina reranking + lesson extraction from full transcripts. Full walkthrough and comparison with the native system: youtu.be/QVVtoaSodHU?si…

Cool to see people building persistent memory for AI agents with LanceDB! This @OpenClaw tutorial by @tonbistudio shows how the community plugin memory-lancedb-pro adds hybrid retrieval, reranking, and session distillation so agents can actually remember across sessions.

@dltHub @huggingface Nice work from the Hugging Face and dltHub teams, congrats on this launch 👏 Check out the blog: dlthub.com/blog/hugging-f…

@dlthub and @huggingface 🤗 just shipped a clean way to ingest Hugging Face datasets into LanceDB 🚀 Query datasets over hf:// with DuckDB, stream them in batches, and load them into LanceDB with embeddings generated during ingest. The result is a simple Python path from Hub dataset to a searchable, explorable table.

6/6 All four layouts share the same on-disk descriptor: (kind, position, size, blob_id, blob_uri) kind determines the storage semantic while the logical column type stays the same. Read more about the mixed blob access problem and the format design behind Blob V2: lancedb.com/blog/lance-blo…

5/6 External (URI): Stores only a URI (optionally with offset/size). Allows existing object storage assets to be referenced directly without copying.

4/6 Dedicated (> 4 MB): Each blob stored in its own file. Isolated from compaction and supports direct object store range reads. Typical for large videos, PDFs, and raw assets.