Pinecone targets agentic completion rates

Pinecone has released Pinecone Nexus, a knowledge engine designed to move reasoning from retrieval to curation. It will stop agents from burning tokens on raw data and start giving them trusted, compiled knowledge. Agents are burning tokens because they are working with systems designed for humans. It causes them to retrieve, read, discard, and, because of missing data, repeat again and again, until eventually they have the answer.

Ash Ashutosh, CEO, Pinecone, said, “We’ve seen AI agents waste 85% of effort on brute-force retrieval.

“Pinecone Nexus shifts reasoning upstream, delivering task-specific artifacts that can cut token costs by 90% and accelerate completion by 30x. This is the future of knowledge infrastructure.”

A problem designed into our technology

Previous technology shifts saw relational databases for client-server and object stores for the cloud. Even vector databases designed for RAG were built around how humans access data, according to Pinecone. The models assume that humans type a query and read the returned documents.

Agents break that because they scan to complete a task. Standard retrieval does not give them what they need. They end up scanning large amounts of irrelevant data and miss connections. That creates hallucinations, which cause problems with trust.

It means we must have a rethink of how data is accessed by agents rather than humans. Most organisations are just at the point of tuning core systems for basic AI access. That now has to change to support both agentic AND human interaction. That’s a complete re-architecture of two approaches that are wildly different.

Can Nexis fix this?

At the core of the problem, according to Pinecone, is a completion rate of just 50-60%. That is compounded by completion times being unpredictable, which makes it hard to set SLAs for agentic projects. Additionally, the soaring token cost and waste of tokens by agents threaten any attempt at ROI for these projects.

Nexus is Pinecone’s solution to changing how we make data accessible to both humans and agents. Unlike previous technologies, where retrieval is the goal, Pinecone says Nexus is a knowledge engine.

What’s the difference? According to Pinecone:

“A retrieval system finds documents and hands them to a frontier model at inference time. The model burns tokens, sifting through raw content, introduces latency, and risks hallucination. This is reasoning at the retrieval stage. It is expensive, slow, and fragile.

“Nexus moves the reasoning upstream, from retrieval to knowledge compilation. It structures, contextualizes, and composes specialized contexts (derived artifacts) before the agent needs them. The agent receives trusted knowledge in a context-specific structured format, not raw documents. It completes the task, not the retrieval. Frontier models are freed to do what they were designed for — intelligent reasoning, not managing knowledge.”

How does Nexus work?

There are two parts to Nexus: a context compiler and a composable retriever.

• The context compiler – takes source data and a task specification. It builds task-optimised artefacts. It experiments with representations and converges on the precise knowledge structure the agent needs. This work happens once at compilation time, not every time the agent runs.
• The composable retriever – serves these curated artefacts at query time: low-latency, grounded, composable across sources. Typed fields. Per-field citations with confidence levels. Deterministic conflict resolution. Output shaped exactly as the agent specified, structured to complete the task accurately and fast.

Pinecone gives an example of a mid-market SaaS company whose data lives in Salesforce, Slack, Gong, and Jira. Using vibe coding to scan and retrieve data does not guarantee the right context surfaces.

The context compiler builds specialised, distinct artefacts for each agent that it uses.

• Sales Agent: Gets deal context. It sees Gong transcripts synthesised with opportunity stages and champion emails. It does not see a CRM lookup. It sees a picture of the deal.
• Finance Agent: Gets revenue context. It links contract terms to billing schedules and usage thresholds. Same Salesforce record, but a completely different artefact.
• Marketing Agent: Gets attribution context. Campaign touches are connected to win/loss themes from Gong and product-qualified signals from usage data. What’s actually driving conversion, not what the CRM says sourced the lead.
• CEO Agent: Gets a cross-functional signal. It links ARR movement to customer health and hiring velocity.

By using four agents, each focused on a specific role, ensures they are optimised for task completion. Humans see data as part of a system of record. The context compiler builds what the agent needs to understand the business.

KnowQL – A new Query Language for Agents

When AI systems first appeared, everyone worried about prompt engineering. Did we need to teach users how to write better prompts? Is this a repeat of the SQL journey? What tools will make this accessible to users?

Enterprises and vendors said users didn’t need to learn prompt engineering. The assumption was that natural language interfaces meant we wouldn’t need to understand prompts. That has changed.

Anthropic’s guidelines for Opus 4.7, say the model “interprets prompts more literally and explicitly than Claude Opus 4.6.” Be precise, do not expect it to read your mind.

OpenAI guidelines for GPT-5.5 says that shorter, outcome-first prompts usually work better than process-heavy prompt stacks. It goes on to say that GPT 5-5 is good at reasoning and figuring out what you want. However, the small print does say you still need to be clear about what you want.

To make it easier for agents, KnowQL adds six core primitives: intent, filter, provenance, output shape, confidence, and budget. This means humans need to think about how they structure what they are asking for. If they get that structure right, the agent will find and return the response they want.

Pinecone is being clever here. It is positioning KnowQL as a generic query interface that should work across all agentic deployments. The question is, will it work everywhere? That is unknown as yet, and there is no sign that it has submitted KnowQL for consideration as a standards-driven language. If it does, it could become the agentic equivalent of SQL.

What we do know is that Pinecone is claiming use of KnowQL will have measurable impacts. It says it will deliver, “up to 90% reduction in token usage. Task completion rates will rise to above 90%, and organisations will see a 30x faster time-to-completion.”

A marketplace for Nexus and new pricing

To drive Nexus adoption, Pinecone has announced 70 production-ready knowledge applications across sales, insurance, legal, and HR. These are fully functioning applications rather than demos built by partners and Pinecone.

Those apps are aimed at four sectors:

• Sales & Revenue: Instant answers on pricing and competitive positioning.
• Insurance: End-to-end apps for underwriting and claims.
• Legal: M&A diligence and employment law tools.
• Customer Support: Frontline Q&A and executive complaint handling.

Importantly, Pinecone says that these apps are ready for immediate deployment with no need to build new infrastructure. All that is required is some configuration.

Nexus also heralds the launch of a new pricing tier. There is a new Builder Tier that offers full access for $20 a month. Dedicated Read Nodes provide fixed pricing for high-volume workloads, reducing costs by up to 97% at scale. For enterprises with data residency requirements, Bring Your Own Cloud (BYOC) is now available.

Enterprise Times: What does this mean?

Companies want to use agentic AI to improve the knowledge flow through their organisation. But to do that in a cost-effective fashion, they need to re-architect their data infrastructure. That’s a hugely costly exercise and one that creates challenges in matching human and agent access.

Pinecone is positioning Nexus as a solution. Rather than re-architect everything, it says Nexus is the agentic knowledge engine. It will enable agents to access data far more effectively than at present. It promises lower costs, faster response times and a better route to ROI. The marketplace and KnowQL will get companies started.

But it does leave questions. How fast will customers move from pilot to production? Which third-party apps on the Marketplace will get the most traction? Is the claim that apps will only need configuration going to prove true? Can we be sure that customers won’t end up writing customer code to overcome issues?

A more important question is, what resources are going to be made available for people to understand KnowQL? The AI query agent on the Pinecone website responded, “I don’t have specific information about dedicated KnowQL training materials in the available sources.”

It continued, “Since Pinecone Nexus and KnowQL are newly announced features in early access, comprehensive training materials may still be in development or available only to early access participants.”

That is disappointing because it suggests that a key element here is going to take time to understand. Until we see more about KnowQL, including who is likely to be using it – IT or users – it’s hard to see how quickly customers will get the best out of Nexus.

The post Pinecone targets agentic completion rates appeared first on Enterprise Times.