Every support ticket used to start from nothing. A customer had a question about a Pinecone account; an AI agent searched the documentation; the agent supplied an answer that sounded plausible and was often wrong. The missing detail might have been the customer’s plan, a recent error, or a change made yesterday. The relevant facts existed in Pinecone’s own systems. Its bot simply did not know which ones belonged together.
That is a deliciously inconvenient problem for a company whose business is helping AI find things. Pinecone began by making vector search available as a managed database. By 2026, it had learned that retrieval was only one chapter of the story. Its own support agent became a test of the next one: could software arrive at a question already knowing enough about the business to answer it?
- Pinecone sells managed search and knowledge infrastructure for AI applications.
- Its database serves semantic, keyword and hybrid retrieval; Nexus prepares company context for AI agents.
- Its own early support trial reported a higher resolution rate, though the newer sample was much smaller.
- Costs begin with a free tier and rise with usage, features and deployment requirements.
The first machine had to find the right neighbor
Founder Edo Liberty had seen the first difficulty while working at AWS. Machine learning turns documents, products, images and behavior into lists of numbers called vectors. Nearby vectors often represent similar things. That makes them useful for recommendations, semantic search and classification. It also creates an unglamorous chore: store millions or billions of those lists, keep them updated, and find the nearest matches quickly enough for a real application.
Pinecone was founded in 2019 and opened its vector database to the public in 2021. The offer was simple: engineers could call an API instead of operating a distributed index. A developer building a search box or a recommendation system still had to decide what to embed and what a good result looked like. Pinecone would look after the machinery that stored and retrieved it. When generative AI made retrieval augmented generation, or RAG, fashionable, that same database became a way to bring private information into a model’s answer.
The work is useful because a model is not a filing cabinet. It can write a fluent response without knowing a company’s actual return policy, contract, or latest product release. A search system gives it something to consult. But “nearby” is a mathematical description, not a promise of truth. A passage about a similar account can be the wrong passage for this account.
The bill comes before the epiphany
Gong offers a useful view of Pinecone’s original promise at scale. Its software studies customer conversations. The company turns sentences into vectors so its Smart Trackers can identify concepts across calls, even when people use different words. Gong says it stores billions of sentence vectors in Pinecone. After moving the workload to Pinecone’s serverless architecture, it reported a tenfold reduction in cost. That is a customer account, not a universal price comparison, but it explains why storage design matters: an AI feature that works beautifully at a thousand records may become extravagant at a billion.
The figures are company and customer reported; they describe different populations and should not be read as audited totals.
Pinecone makes its money from cloud software. In September 2026, its published plans ran from a free Starter tier to a $20 monthly Builder tier, a Standard tier with a $50 monthly usage minimum, and an Enterprise tier with a $500 monthly usage minimum. Database operations, inference and assistants can add usage charges. The more consequential cost can be engineers’ time spent making retrieval useful and keeping it useful when the data changes.
The first thing to fail was context
Pinecone’s support team had a bot that searched its documentation. In the second quarter of 2026, it resolved 24.6% of the tickets assigned to it, according to the company. The rest generally reached a person. Pinecone says staff often needed to look up information that was already in the company’s tables: account plan, recent errors, and recent changes. A generic documentation answer could not substitute for that context.
“Every support ticket used to start from nothing.”Pinecone on its earlier support agent
In July, Pinecone put its new product, Nexus, behind the support agent. Nexus is meant to collect dispersed enterprise sources, prepare structured knowledge from them, and give an agent something more useful than a handful of similar passages. Pinecone reports that the new agent resolved 55.1% of tickets assigned to it in the early period after the change. The denominator was 49 assigned tickets; the earlier rate came from 403. The result is promising and specific, though too early to treat as a general law of support automation.
The lesson is portable. Start with a narrow workflow where a person repeatedly assembles the same context. List the facts they check before answering. Connect those facts to the agent, keep them fresh, and measure the outcome that matters to the customer rather than just the agent’s speed. A retrieved document is a clue. A resolved problem is the result.
Why exact words came back
Pinecone has also broadened its database in a revealing direction. In September 2026 it made full text search generally available alongside vector search. A semantic index is good at finding passages with similar meaning. It can be poor at a precise part number, error code, order ID or legal citation. Those strings do not ask to be understood; they ask to be matched. Pinecone now puts keyword ranking and text filters in the same index as dense vectors, so an application can use both kinds of evidence.
This is also where Pinecone sits in a crowded market. Dedicated rivals include Qdrant, Weaviate and Milvus. Teams can add vector search to PostgreSQL, OpenSearch or other systems they already run. Pinecone’s case is strongest when a team wants a managed retrieval service, developer friendly APIs and scale without owning index operations. For a small dataset already living in PostgreSQL, adding another service can be harder to justify. Nexus then asks a different buying question: how much repeated agent work can be avoided by preparing knowledge ahead of time?
A company selling fewer blank slates
Pinecone has followed that argument up the software stack. Database retrieves records. Inference hosts embedding and reranking models. Assistant helps developers build document grounded answers. Nexus prepares enterprise context for agents, and its bring your own cloud option places the data plane in a customer’s AWS, Google Cloud or Azure account. The company says its mission is to “make AI knowledgeable.” That phrase sounds grand until one remembers the support ticket: the useful knowledge was simply which account was in trouble and what had happened to it yesterday.
The leadership change fits the widening brief. Liberty, a scientist and former AWS research leader, passed the CEO job to Ash Ashutosh in 2025 and moved toward research. In 2026 he stepped back from daily work while remaining on the board. Ashutosh, a veteran of storage and cloud businesses, inherited a database company entering a more complicated market. Pinecone says more than a million developers and 10,000 customers use its foundation. The next test is whether those customers will pay Pinecone to organize the knowledge their agents need, not merely fetch the closest text.
That distinction may be the whole profile. The old bot could read. The new one had to remember. A business that once sold nearest neighbors is now trying to make the right neighbor show up with a name, a date and the customer’s account attached.