Pinecone

Use TypeSafe AI's Jev model with Pinecone to rerank results with natural language criteria! Usually with rerankers, it's hard to cleanly specify what should and shouldn't be returned in results. Jev resolves this by refactoring the problem into evaluating against distinct binary criteria, which pairs great with Pinecone retrieval! In this demo, we compare using Jev and Claude to rerank 200 returned candidates from Pinecone. Jev returns a reranked list in about a second — 830 to 1,300 ms across eight test queries. Claude Opus 5, doing the same job in one long-context call, takes 4.2 to 6.8 seconds. That's roughly 5x faster, and about 43x cheaper: $0.004 per query against $0.18. Try it out, here!

A new release of our Agent Skills and Docs eval library, Cultivar is available! Now, use TypeSafe AI's Jev model to evaluate how well your agents use docs and Skill against specified tasks in Modal sandboxes. This works great, because Jev allows for binary yes/no decisions calibrated with respect to criteria, exactly the format Cultivar works best with! Preliminary results benchmarking grading agent runs with Jev over Claude produces a 19x speedup and is 38x cheaper! However, using Claude allows for detailed reasoning and justification for each run. Use Claude for complex evaluations and Jev for quick ones. Install and use cultivar here: https://lnkd.in/gNGQFSkk

Vector quantization is fundamental to vector databases. Before you can search vectors, you have to store them, and storing them at full precision is expensive. Every bit you save shows up as capacity and cost. Pinecone has used quantization since its first prototypes, and we keep looking for the state of the art. So this summer we set out to survey it. We found more papers than we expected, and no way to compare them: every paper measured different metrics, on different datasets, tuned for different hardware. We could not find a single systematic evaluation of the leading methods against each other. What we did find was a pattern. Most published quantizers are built from the same small set of primitive operations, composed in a different order. So we built VQ-bench: an open-source library of those primitives, composable into pipelines, and designed to be extended. E-RaBitQ, one of the strongest methods we tested, is four primitives in a list. Swap one and you have a new quantizer. Add a primitive of your own and it composes with everything already in the catalog, including pipelines nobody has written yet. Either way you can measure the result against every other method with the same harness. We used it to benchmark 14 popular quantizers on recall, reconstruction error, and encode time. Two takeaways so far: PQ and OPQ have the lowest reconstruction error, and EDEN keeps up with E-RaBitQ on recall at higher bit budgets while encoding much faster. This is a first iteration. We want your feedback, corrections, and contributions, and we will keep adding quantizers over time. Website: https://lnkd.in/gPufqSGF Blog: https://lnkd.in/gVb9zy8D Repo: https://lnkd.in/g53PEuyD Paper (VecDB@VLDB 2026): https://lnkd.in/gpiX7wDW

Congratulations to Matan Wittner, our new Site Reliability Engineer, who brings a wealth of experience in cloud infrastructure and security. We are excited to have you on the team! Interested in joining our talented team at Pinecone? Check out our opportunities today: https://lnkd.in/d4R_7sN

A database knows where your data is. A vector database understands what your data means. Instead of relying on exact keyword matches, it turns text, images, audio, and video into vectors so it can find things based on similarity and meaning. That’s what powers things like semantic search, recommendations, RAG, and agent memory. Watch the video, then learn more about how vector databases power AI applications: https://lnkd.in/gbSBJe3K

Full-text search (FTS) became generally available in Pinecone Database earlier this week. It was one of our most requested features, and it runs alongside vector search in the same index. Pinecone queries can now combine full-text and semantic search in a single request. Match exactly on the terms that have to be right, then rank what's left by meaning. Say you're building a bird identification agent. You ask it for woodpeckers that look "red crested." The full-text match narrows the set to woodpeckers, and the image embeddings rank those by how closely they match a red crest. What you get: - Exact matches on the things that have to be exact: IDs, SKUs, error codes, product names - BM25 keyword ranking across multiple text fields, with Lucene query syntax, boolean logic, and fuzzy matching - One index for text, dense, and sparse, with no separate search engine to run - Usage-based capacity by default, with Dedicated Read Nodes and BYOC when you need them Thanks to everyone that gave us feedback during our preview. Read our blog post and try it out. 🔗 https://lnkd.in/giKfwNc8

At Pinecone, we’re seeing BYOC adoption is growing for different reasons: data sovereignty, security requirements, control, and keeping data close to where it already lives. The interesting engineering challenge is how to give customers that control without giving up the simplicity of a managed service. Looking forward to discussing exactly this with Eran Kampf David Hayes and Jennifer Riggins at LeadDev. https://lnkd.in/gnABKtJA

The LA Agentic AI Meetup is on the official Tech Week calendar this year. We'll be at Gulp in Playa Vista on Tuesday, October 13, from 5 to 7pm. We've been running this meetup since May, so expect the regulars plus everyone who shows out for Tech Week. It's a low-key crowd. Networking, drinks, a few live demos, and real people comparing notes on what they're building and where things are going. Whether you're deep into RAG and agentic workflows or just starting to explore AI, you'll find people doing the same. RSVP: https://lnkd.in/gguh9qve TECH WEEK by a16z

Leading this project was the hardest thing I’ve done in my career. It dragged me out of my comfort zone and into work I didn’t expect to do. Software engineering is a craft you can only learn by doing, and nothing teaches you faster than being thrown in at the deep end. Looking back, some of my most useful contributions weren’t technical. My focus shifted from writing code to evaluating tradeoffs, making difficult decisions, setting the roadmap, communicating plans, and keeping the team aligned as the work evolved. With AI, code has become cheaper to produce, but context, judgement and leadership are more valuable than ever. In addition to indexing text for search, we added a document data model that supports flexible fields and types, a schema to describe how documents should be indexed, and powerful new APIs. All of these features were threaded through a mature database without disrupting a single customer in production. The result is a search engine that can find exact matches using full-text search, similar items using vector search, or combine both approaches in a single query to find the most relevant results. Whether you need billions of documents, millions of namespaces, thousands of requests per second, or millisecond response times, Pinecone can handle your most demanding search workloads. Thanks to everyone who got this over the line: Rajat Tripathi, Austin DeNoble, Ivan Kelly, Clint Wang, Yaakov Sokolik, Jennifer Hamon, Nico Alba, Jenna Pederson, Arjun Patel, Manish Talreja, Keith Corbalis, Chris Bolton, Jörg Schad , Ph.D.