Pinecone slashes time required for developers to build AI solutions

Pinecone has integrated inference capabilities into its vector database. This, it says, will create AI applications faster and be up to 48% more accurate. The company is achieving this by adding several new technologies to its database solution, which it claims will define a new standard for AI-powered solutions.

Edo Liberty, founder and CEO of Pinecone, said, “Our goal at Pinecone has always been to make it as easy as possible for developers to build production-ready knowledgeable AI applications quickly and at scale.

“By adding built-in and fully-managed inference capabilities directly into our vector database, as well as new retrieval functionality, we’re not only simplifying the development process but also dramatically improving the performance and accuracy of AI-powered solutions.” 

What is Pinecone adding to its composable platform?

The company has listed five updates to its platform. They are:

• pinecone-rerank-v0 proprietary reranking model
• pinecone-sparse-english-v0 proprietary sparse embedding model
• New sparse vector index type
• Integration of Cohere’s Rerank 3.5 model
• New security features, including role-based access controls (RBAC), audit logs, customer-managed encryption keys (CMEK), and the general availability (GA) of Private Endpoints for AWS PrivateLink

Speeding up retrieval

At the core of these updates is the need to improve the ability to retrieve data. Pinecone is addressing this through two approaches. The first is to rerank models using a new proprietary approach.

Reranking works fastest when it is implemented as a two-stage process. First a set of documents is retrieved that meet the main search criteria. That set of documents is then reranked based on a second set of criteria.

In addition to reranking, Pinecone has also added dense vector retrieval. It is a technique where semantically close texts are represented by vectors in a close vector space. It is especially effective when working with unstructured data.

Another technology is sparse retrieval. It focuses on the frequency distribution of terms, making it possible to precisely locate keywords. Pinecone says this makes it a key approach for domain-specific terminology. For those working with specialised documents such as legal, healthcare, and scientific data, this provides a significant increase in data retrieval.

According to the announcement, “combining the sparse retrieval, dense retrieval, and reranking capabilities within Pinecone, developers will be able to create end-to-end retrieval systems that deliver up to 48% and on average 24% better performance than dense or sparse retrieval alone.”

Introducing two proprietary models

As part of this release, Pinecone has introduced two proprietary models. It claims they will make it “easier for developers to build knowledgeable AI.” The two models are:

• pinecone-rerank-v0 improves search accuracy by up to 60% and on average 9% over industry-leading models on the Benchmarking-IR (BEIR) benchmark
• pinecone-sparse-english-v0 boosts performance for keyword-based queries, delivering up to 44% and on average 23% better normalized discounted cumulative gain (NDCG)@10 than BM25 on Text REtrieval Conference (TREC) Deep Learning Tracks

Support for Cohere Rerank

In addition to its own technology in this release, the company has collaborated with Cohere. It will host cohere-rerank-v3.5 natively within the Pinecone platform. For customers, this allows them to choose the Cohere-rerank-v3.5 directly from the Pinecone API to enhance the relevance of their search results.

According to the announcement, “Rerank 3.5 excels at understanding complex business information across languages making it optimal for global organizations in sectors like finance, healthcare, the public sector, and more. By incorporating Cohere’s latest industry-leading reranking model, developers can further refine search outputs, ensuring more accurate and contextually relevant responses for their applications.”

New security controls on Pinecone’s database

As part of this release, Pinecone has also boosted the security controls around its database. It is a good move and the first of the improvements allows users to set API key roles for access is important. With databases being increasingly accessed via APIs, organisations need to review the level of access those APIs have. When those databases are part of AI projects, this is even more important.

The four improvements that the company has listed are:

• More granular role-based access controls (RBAC) lets users set API key roles for control and data plane operations
• Customer-managed encryption keys (CMEK) enable users to control their own data encryption and enhance tenant isolation
• Audit logs for control plane activities (e.g. index creation or deletion) via Amazon Simple Storage Service (Amazon S3) endpoints
• Support for AWS PrivateLink is now generally available (GA) for serverless indexes

Enterprise Times: What does this mean?

This is an interesting announcement from Pinecone. The focus on how to improve the retrieval of data to speed up AI is something other vendors are likely to do. Adding its own technology to improve reranking and sparse retrieval shows the commitment to improving the product. It also shows how the company is looking to differentiate itself from other vendors.

Will it deliver the speeds it claims? Some early customer comments suggest it is more than achieving that. The question is, are they outliers or will the vast majority of customers get that level of performance improvement? That opens another question which is, what will Pinecone do next? How much more performance can it get out of its database?

Of course, there is always a concern when improvements are down to proprietary solutions. Will customers find themselves feeling locked in? How will they move to a competitor? Given the size of the gains, it is unlikely that many will worry about those concerns.

So what next for Pinecone? It last did a funding round in April 2023 when it raised $100 million at a valuation of around $750 million. Any new funding round would almost certainly make it a tech unicorn. Will this latest product announcement, is it finally in a place where it can bootstrap all future development and even consider a IPO?