Pinecone

Your coding agent doesn't need to leave the terminal to use Pinecone. We now ship official plugins and skills for the agentic IDEs and CLIs you are already building in: Claude Code, Cursor, GitHub Copilot, Codex, and any MCP-compatible client. Every tool gets the same skill set, including two new ones. Use the full-text-search skill to create, ingest into, and query a document-based index. Use the n8n skill to build n8n workflows with Pinecone Assistant or database. All without leaving your agent. These sit alongside skills for on-boarding to Pinecone, querying, building with Assistant, and CLI workflows. Install the skills for your agent here: https://lnkd.in/g8m8sZrY

Semantic search finds what's close in meaning, not what you actually meant. Ask "who are the top presidential candidates?" and a vector search can hand back three results about the French election (accurate, but for the wrong election). A person scanning through catches that in a second and skips past it. An agent doesn't get that chance: it treats retrieval as ground truth and starts acting on it. Pinecone's text match filters narrow the candidate pool with a lexical constraint before the vector search runs, in order to get the right context without pre-labeling a dataset for every dimension a query might need. Read how it works, with a real example: https://lnkd.in/gZSrqApf

4 pricing lessons every AI infra company can learn from Pinecone: Everyone says pricing vector databases is hard. I agree. Because you're not pricing a seat or a flat SaaS tier. You're pricing storage, reads, writes, embeddings, reranking, and an assistant layer that all moving at different speeds. Pinecone just rebuilt their pricing around this reality. Here are 4 moves AI infra companies can steal from them: 1. A real free tier that maps to actual workloads Instead of vague "free credits," they show you what the Starter plan actually buys: ~15K semantic searches/day, ~44K recommendations/day, ~130K RAG chats/day. They translate limits into outcomes. Developers instantly know if they fit. 2. A new $20 "Builder" tier to catch the missing middle Between free and a $50 usage-based plan sat a gap: the solo dev who's past the free tier but not in production. The flat $20 Builder plan (multiple projects, choose your cloud, monitoring) fills it and quietly moves people from "just testing" to "committed." 3. Charge on cost drivers, not a made-up metric Storage ($/GB), Write Units, Read Units, egress, tokens, ingestion units. Every dimension maps to something that actually costs Pinecone money. When your pricing mirrors your COGS, margins don't surprise you at scale. 4. Minimums that pre-qualify seriousness $50/mo min on Standard. $500/mo min on Enterprise. It's not just revenue but it's a filter. The minimum tells a buyer which plan is "for people like us" before they read a single line item. The thing I'd watch: 6+ metered dimensions is transparent, but it's also a lot to forecast. A developer can't easily answer "what will this cost me next month?" without the calculator. Transparency and predictability aren't the same thing.

Sparse indexes handle keyword and lexical search: each term maps to a posting list of every document that contains it, scored by BM25 for exact term matches or SPLADE for transformer-expanded matches that catch related terms too. We just rewrote how that posting data gets stored on disk. The previous version stored posting data by document range, so a query touched every block in the index regardless of which terms it referenced. That held up while indexes fit in memory. Once sparse indexes grew past memory limits, every query became a full disk scan, and latency stopped depending on the query and started depending entirely on disk throughput. V3 reorganizes the index by term instead. Each term owns its own blocks on disk, and a query loads only the blocks for the terms it contains. BM25 queries now read 1,428x less data and return in single-digit milliseconds. SPLADE queries read 151x less and run 27x faster. Recall holds for SPLADE and improves for BM25. Full writeup: https://lnkd.in/gTf-3Nyd

Random vectors work fine for testing write throughput. Recall and accuracy tests need something closer to real data. A Pinecone Forward Deployed Engineer built a repeatable pipeline for generating Pinecone test data: real news articles from Hugging Face's CC News dataset, chunked and embedded with BAAI/bge-large-en-v1.5 (1024 dims), landed in Parquet, then bulk imported. Small tests can stream, embed, and upsert in one process. Past a few million vectors, the embedding step bottlenecks that. So the pipeline splits: text-only Parquet first, embeddings generated offline, then import. That's the difference between testing at 500K vectors and testing at 100M. 🔗 https://lnkd.in/griHJSia

We're hosting the LA Agentic AI Meetup this Thursday, July 9, 5 to 7pm at Gulp in Playa Vista. Come talk to engineers, founders, and builders working across RAG, agentic workflows, and the broader AI stack. Drinks and real conversations about what people are shipping. New this time: two live demos. Michael Campbell, PhD, is showing an agentic job-search system, and Andre Calloway-Cazares is demoing Fresh House, a Carfax-style condition record for homes. RSVP: https://lnkd.in/gYMKfPve

Happy Independence Day America! 🇺🇸 250 years ago, our founding fathers shook free from the oppressive tax regime of the British monarchy. Today, we are faced with another tax: the token tax on retrieval when running agents. Over 85% of your token spend is going towards re-retrieving the same data and business context every time your agents start executing a task. That sounds like a tax that no one wants. We built Pinecone Nexus to break you free from this new oppressive token burning regime. Instead of having agents running in circles retrieving data from your contracts, wikis, HR documents, meeting notes, support tickets, financial records, etc every single time, Nexus compiles everything upfront into knowledge artifacts. All queries after that are faster, cheaper, and more accurate. Our customers are seeing more than 90% in token savings. That sounds like freedom to me. This 4th of July, add our new public preview to your hot dogs and fireworks celebrations. Try it out: https://lnkd.in/ggVsUSUr Happy Independence Day. 🎆

"Essentially, Nexus provides a framework that explains the underlying 'where to find what you’re looking for' for the AI agent. This is similar to how an employee with over three years at the company knows which part of the financial archives to consult for acquisition data or where to look for the most recent business logic for the web portal. "This is different from prompt engineering, where users or engineers must teach the agent where to look at query time... "In benchmarks, Pinecone said Nexus demonstrated high performance and accuracy. "'We can easily stand up a vector database and run RAG (and agentic search) over our documentation corpus,” Jesse Barbour, Chief Data Scientist of Q2 Holdings Inc., an Austin-based financial technology solutions company. 'The hard part is getting an agent to reliably and efficiently assemble the right knowledge for genuinely difficult questions.' "According to Barbour, Nexus answered complex support questions with 95% accuracy. It also kept token costs low, making it an enticing knowledge layer as AI inference prices rise." — Kyt Dotson for SiliconANGLE & theCUBE 📰 https://lnkd.in/gcHSkAjX

Congratulations to Jackson Gold, our new Associate Field Engineer, who comes to us with a wealth of experience in AI and startups. We're 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