Pinecone’s cover photo
Pinecone

Pinecone

Software Development

New York, NY 71,993 followers

Build knowledgeable AI

About us

Pinecone is the leading vector database for building accurate and performant AI applications at scale in production. Pinecone's mission is to make AI knowledgeable. More than 5000 customers across various industries have shipped AI applications faster and more confidently with Pinecone's developer-friendly technology. Pinecone is based in New York and raised $138M in funding from Andreessen Horowitz, ICONIQ, Menlo Ventures, and Wing Venture Capital. For more information, visit pinecone.io.

Website
https://www.pinecone.io/
Industry
Software Development
Company size
51-200 employees
Headquarters
New York, NY
Type
Privately Held
Founded
2019

Locations

Employees at Pinecone

Updates

  • If you have questions about how to get the most for your 💲 with Pinecone, let us know. Our team, including Head of Field Engineering Perry Krug, will let you know how to get the best cost-performance for for your workloads in the industry! Here's the DRN announcement Perry references 👇 https://lnkd.in/gJHHPwh3

    Had a customer recently migrate from our pod-based indexes to our serverless architecture and went from paying a flat $286/day regardless of their traffic to ~$32/day on weekends and ~$200/day during the week, on track to save them over $1k/month even after doubling their traffic month over month A few points I called out: - They are now paying MUCH less for their non-production workloads rather than having them sit in pod indexes that were mostly idle. For instance, all of their dev indexes combined cost them less than $1 last week on serverless whereas each one alone on pods would have been at least $4 for the week. - Another important consideration here is the alignment of "unit economics" between their users and their cost on Pinecone: their larger and more active users will cost them more, but their smaller ones will cost them less. For example, they have around 145 namespaces on Pinecone. The top one has around 19M records in it and costs them 124 RUs ($0.003) per query, and received around 430k queries in the past month (about $1200). At the other end, one of their smallest namespaces only has 331 records in it which costs them .26 RUs ($0.000006) per query and only received 5 queries in the past month. They have many more of the small ones than the large ones. In a pod-based index, or many other technologies, not only would they be paying to have that data sitting idle but they would also have to deal with the operational overhead of sizing and re-sizing your namespaces (i.e. their customers/tenants) as they grow and change. That doesn't exist with Pinecone's serverless offering. - To some degree, this is the nature of a usage-based system vs an hourly one...on pods they were spending money by the hour regardless of whether they were using the system or not. With serverless, the expectation should be that although they may pay more for a given day of high usage, that will be offset by the times they're paying less or nothing at all. And that the cost grows in line with the success of their business. - For cases where it *DOES* still make sense to pay by the hour and amortise the cost of each read, Pinecone's Dedicated Read Nodes are a perfect fit. The more queries you do, the less each costs and you can scale the infrastructure up and down as needed.

  • ICYMI: We launched bulk data operations by metadata this year 🎯 Update, delete, and fetch records using metadata filters – no ID collection needed. If you're already using metadata filters in queries, you know the syntax. Same filters, more power. This was one of our most requested features, and with the latest API updates, even more is now available. Catch up on what you missed: https://lnkd.in/gXCbGHrB

  • Pinecone reposted this

    We made hamming distance scans ~30% faster. I wrote the first C/C++ implementation of hamming similarity approximation for v0 Pinecone in 2019. It was pretty fast. A couple of years later we (finally) replaced it with Rust. That was faster. The Rust compiler optimized the SIMD instructions better than I did. We still use hamming distance scans in places and we thought it could not get any faster. We were wrong. We recently revisited that code to explicitly invoke AVX-512 intrinsics with loop unrolling. It’s now ~30% faster… Bits | Previous (ns) | Unrolled AVX512 (ns) | Throughput Improvement 512 | 2.02 | 1.49 | 1.35× 1024 | 3.04 | 2.11 | 1.44× 2048 | 5.64 | 4.48 | 1.26×

  • ICYMI: Dedicated Read Nodes are now in Public Preview Need predictable performance for high-throughput vector workloads? Pinecone's newest feature gives you reserved capacity with: ✓ Hourly per-node pricing for cost predictability  ✓ Warm data paths (memory + SSD) for consistent low latency  ✓ Linear scaling: add replicas for throughput, shards for storage Perfect for billion-vector semantic search, recommendation systems, and mission-critical AI services with strict SLOs. Choose what fits your workload: → On-Demand for bursty, elastic traffic → DRN for always-on, high-QPS applications Learn more about Dedicated Read Nodes: https://lnkd.in/g8b5WCTm

    • No alternative text description for this image
  • Standing room only at Arjun Patel's lightning talk yesterday at re:Invent! His take? We need Agentic RAG for complex queries that require real reasoning and planning. The example that hit home: trying to find a restaurant for a couple with totally conflicting dietary needs and prefernces (traditional RAG just can't handle it). The solution is agents that can decompose queries, dynamically pull from multiple sources, and iterate on outputs before returning results.   

  • Pinecone reposted this

    The energy at Amazon Web Services (AWS) re:Invent is electric this year! It’s inspiring to spend time with the builders of the future: customers pushing the boundaries of AI, partners bringing innovative ideas to life, and developers turning ambitious concepts into real-world systems. I am incredibly proud to be here with the Pinecone team. In just a few months, I’ve seen the team's deep commitment to solving real problems for builders. This week is particularly significant for us. We are announcing Dedicated Read Nodes (DRN) in Public Preview! This represents a major step forward, fundamentally changing the economics of AI and Agentic apps at scale. DRN is a new deployment mode for our unique serverless “slab” architecture, optimized for workloads involving millions to multi-billions of vectors and predictable query rates. By combining DRN with our existing on-demand options, we ensure customers achieve the optimal cost-performance balance for any AI use case—from RAG and agents to search and recommenders. The next wave of knowledgeable AI is here, and Pinecone is purpose-built to support it. I look forward to the conversations ahead this week and all that we will continue to build with our community. See the full details on the DRN announcement here. 👇 #AWSreInvent #Pinecone #VectorDatabase #AI #GenerativeAI #RAG #AgenticAI Dedicated Read Nodes Announcement blog: https://lnkd.in/e72WDWh5

    • No alternative text description for this image

Similar pages

Browse jobs

Funding

Pinecone 4 total rounds

Last Round

Secondary market
See more info on crunchbase