Getty Images/iStockphoto

Pinecone unveils serverless vector database, targets costs

The vendor's new tool is designed to help customers control their cloud computing costs while developing and maintaining generative AI applications and models.

Pinecone unveiled a new serverless vector database designed to reduce the cost of infrastructure management while helping improve the accuracy of generative AI applications while.

Pinecone Serverless is now in public preview in AWS cloud regions and plans are in place to make the new database available on Microsoft Azure and the Google Cloud Platform next.

Based in New York City, Pinecone is a vector database specialist whose capabilities enable users to store, discover and operationalize unstructured data to train applications and models used to inform business decisions. To date, the vendor has raised $138 million, including $100 million in April 2023 and $28 million in March 2022.

Vectors are numerical representations of unstructured data, such as text and video, that give the unstructured data a value, or structure, so it can be searched and discovered. Vector databases, meanwhile, date back to the early 2000s.

Historically, vectors were used by search-driven organizations that collected huge amounts of data. Vectors and vector databases enable similarity searches much like graph databases, allowing users to discover relevant data amid potentially billions of other data points.

While a niche product for most of two decades, vector databases have skyrocketed in popularity over the past year. Organizations have recognized their value in helping train generative AI large language models (LLMs), such as ChatGPT and Google Bard, with proprietary data so they can understand a given enterprise's operations and deliver accurate results.

In particular, vector databases have gained popularity because they can help train retrieval-augmented generation (RAG) pipelines, according to Doug Henschen, an analyst at Constellation Research.

"Vector embeddings help data scientists with RAG to improve the accuracy and relevance of LLMs based on an organization's own data," he said. "Developers can then deploy these more accurate models as well as more accurate search capabilities within the context of company- and industry-specific applications to drive better results."

Donald Farmer, founder and principal of TreeHive Strategy, likewise noted the important role vector databases play in generative AI development.

"Vector databases are critical for providing the large volumes of vectorized text and data that LLMs need," he said. "By efficiently storing and indexing vector data, they enable significant improvements to AI application quality, reducing hallucinations and enabling LLMs to generate outputs, which adhere to business rules for a specific scenario."

In addition to Pinecone, vector database specialists include Milvus and Chroma, among others. Beyond specialists, other database and data platform providers that now offer vector databases include MongoDB and Snowflake.

New capabilities

The cost of cloud computing has caught many organizations off guard.

Over the past decade or more, the cloud has become the de facto environment for most data management and analytics operations. Meanwhile, the prices for using cloud-based platforms are seemingly low at just a few cents per user per minute or a few dollars for a given amount of compute power.

But given the volume of data organizations now collect and use to make decisions and the number of users spending large amounts of time using cloud-based tools, the cost of using the myriad cloud platforms that make up an organization's data infrastructure has, in many cases, far exceeded expectations.

As a result, organizations are trying to better control cloud computing costs. Vendors are attempting to provide efficient tools that help customers predict and limit usage and spending.

One of those ways vendors are providing more efficient tools is with a serverless computing architecture. Serverless cloud services automatically scale up or down without administrators having to commit in advance to using specific amounts of compute power and storage capacity, Henschen noted.

"Serverless is now a widely used approach across public clouds promoting ease of administration and elastic use of resources to save money while meeting performance requirements," he said.

Pinecone's introduction of Pinecone Serverless on Jan. 16, therefore, has the potential to benefit both existing customers looking to lower costs as well as attract new users seeking cost control, he continued.

"Pinecone is a startup, so it doesn't have a huge customer base at this point," Henschen said. "This will give existing customers the option to switch to the serverless cloud service. But the company's broader hope is surely to attract more new customers that might have balked at having to set up, administer and scale a database using the old-school, manual approach."

A chart displays the differences between keyword search and vector search.
A comparison of vector search and keyword search.

Pricing for Pinecone's traditional database is based on usage, which includes storage and read and write units. The vendor does not disclose what it charges for use but provides a cost calculator on its website.

The vendor, meanwhile, claims that its new serverless database has the potential to result in significant cost savings compared with using databases that require back-end infrastructure management.

Public preview pricing for Pinecone Serverless is 33 cents per gigabyte, per month for storage; $8.25 per million read units; and $2 per million write units.

Like Henschen, Farmer noted that Pinecone Serverless is an important new tool for Pinecone users.

"Serverless is a significant announcement," he said. "I have heard a lot of demand for this over the last year. It provides a serverless vector database architecture that should scale to handle huge amounts of vector data at lower costs. This should enable more companies to improve their AI applications by adding almost unlimited knowledge to GenAI applications."

Farmer added that Pinecone Serverless might be an industry-first.

"Their new architecture … seems innovative in the vector database space," he said.

Future plans

With Pinecone Serverless now in public preview, Farmer noted that the new database can still be improved before becoming generally available and after that with updates.

In particular, Pinecone would be wise to continue focusing on helping customers lower costs, he said.

Meanwhile, adding new users through integrations with partners and tailoring its database to meet the needs of customers in different industry verticals are opportunities for growth.

This will give existing customers the option to switch to the serverless cloud service. But the company's broader hope is surely to attract more new customers that might have balked at having to set up, administer and scale a database using the old-school, manual approach.
Doug HenschenAnalyst, Constellation Research

"I'd like to see Pinecone continue optimizing its vector database for performance and cost efficiency as adoption continues accelerating," Farmer said. "Expanding integrations with other AI/ML platforms could also help drive further adoption. Industry-specific tailoring of the database could be another potential area to explore."

Henschen, meanwhile, noted that while vector search and storage capabilities are gaining popularity given their enablement of RAG pipelines and generative AI development, Pinecone and other dedicated vector database vendors might face stiff competition from other database vendors.

But the specialists position themselves as offering products that are better suited to the needs of users. These come with capabilities that differentiate them from the vector databases developed by broader-based data management vendors, such as SingleStore and Databricks, he said.

Whether those capabilities provide enough differentiation to convince an enterprise to add a new product to their data infrastructure, especially when they can potentially add something similar from a provider they already use for other needs, remains to be seen.

"Pinecone and other dedicated vector database providers make the case that their products are a better fit for AI developers and data scientists with more bells and whistles for their specific needs," Henschen said. "Time will tell if these constituents will see the need to manage and maintain a separate product or pay for an additional database service that's only used for AI development."

Eric Avidon is a senior news writer for TechTarget Editorial and a journalist with more than 25 years of experience. He covers analytics and data management.

Dig Deeper on Data management strategies

Business Analytics
Content Management