Pinecone Changelog: Hybrid search in one index, BYOC, and VQ-Bench

Video authored by the channel shown, not by IANIA; all rights belong to the creator. Transcript and summary generated automatically; the original-language content is preserved.

Original language: English

Original video: Pinecone Changelog: Hybrid search in one index, BYOC, and VQ-Bench · Channel: Pinecone · 03:35 · Watch on YouTube

Topics: Data, databases and embeddings, AI infrastructure and hardware, RAG, search and information retrieval

Summary

This changelog covers Pinecone updates for hybrid search, BYOC, vector quantization, and cost reporting. The Documents API now supports integrated inference and lets one index handle dense, sparse, full-text, and metadata search workloads. BYOC is generally available on AWS, GCP, and Azure, with on-demand capacity support. Pinecone also released VQ Bench, an open-source project for evaluating vector quantizers, and a cost and usage export API for enterprise customers.

Key points

The Documents API supports documents with schemas for multiple distinct searchable fields.
Integrated inference lets Pinecone embed and search data through the same API.
A single index can support dense, sparse, full-text, and metadata search workloads.
Hosted embedding models for dense text fields can eliminate a separate embedding API call.
BYOC is generally available on AWS, GCP, and Azure.
With BYOC, the data plane runs in the customer’s cloud while the control plane remains with Pinecone.
BYOC supports on-demand capacity, so capacity can grow as demand increases.
VQ Bench is a paper, open-source library, and benchmark for building and comparing vector quantizers.
The cost and usage export API gives enterprise customers up to a year of data in the FinOps Open Cost and Usage Specification format.

Chapters

Start	Chapter	Summary
00:00	Documents API and hybrid search	Pinecone introduces integrated inference in the Documents API and explains how one index can support multiple search methods.
01:32	BYOC and on-demand capacity	BYOC is generally available across three cloud providers and now supports on-demand capacity.
02:34	VQ Bench	Pinecone presents an open-source library and benchmark for building, evaluating, and comparing vector quantizers.
03:05	Cost and usage export API	The new API lets enterprise customers export up to a year of cost and usage data in a standard format.

Quotes

you can have Pinecone embed and search all of your data in the same API. — 00:30

Now, you can hit the same index for all of these workloads. — 01:02

As your demand grows, your capacity can grow with it. — 02:02

Quantizers are algorithms that can compress vectors into smaller data representations. — 02:34

We’ve also released a cost and usage export API for enterprise customers. — 03:05

Concepts and entities

Pinecone, Documents API, hybrid search, BYOC, VQ-bench, vector quantization, Cost and Usage Export API, FOCUS

Full transcript (original language)

00:00 Easier hybrid search, the documents API, BYOC, and quantizers. The Pinecone team has been absolutely cooking, and in this video, we’ll recap those changes and more. >> [music] >> Hi, this is Arjun from Pinecone, and this is the Pinecone changelog. Documents API, now with integrated inference. >> [music] >> Implement hybrid search across dense embeddings, sparse embeddings, metadata, and traditional full-text search more easily with the documents API. Starting today, you can use integrated inference with the documents API, which means you

00:30 can have Pinecone embed and search all of your data in the same API. Documents API is the new standard way of making indexes in [music] Pinecone. Before, we had vectors and records, which supported vectors and text respectively. [music] The documents API instead lets you specify a document, which has a schema specifying multiple distinct searchable fields. One purpose of the documents API is to make hybrid search workloads a lot easier with one index supporting multiple data types and searches. >> [music] >> Previously, to implement hybrid search,

01:02 that kind of search that combines one or more search methods like sparse, dense, or full-text, you needed a Pinecone index for each one. Now, you can hit the same index for all of these workloads. When we launched this, you could have a vector field, which is useful for a dense vector, a sparse field for sparse vectors, full-text search fields for typical FTS needs, and metadata. But today, you can have your dense text field [music] use our hosted embedding models instead, which is one less API call for you to hit, and a significant speedup compared to hitting the

01:32 embedding API and searching separately within Pinecone. We’re focused on making hybrid search even easier to use on Pinecone, so stay tuned for more announcements in this area soon. >> [music] >> BYOC and on-demand. We also recently made bring your own cloud generally available, an often requested enterprise feature, alongside bringing on-demand capabilities to it as well. BYOC is available on AWS, GCP, and Azure. It lets you keep your customer data and knowledge in your accounts while Pinecone manages platform ops,

02:02 which is great for financial, legal, healthcare, and other sensitive enterprise workloads. Specifically, the data plane portion of Pinecone runs inside your cloud, hence BYOC, [music] and the control plane remains with Pinecone. Now, BYOC also supports our on-demand capacity mode, which means you don’t have to figure out your index sizing in advance. [music] As your demand grows, your capacity can grow with it. >> [music] >> VQ Bench. We released VQ Bench, a paper, an open-source library, and a benchmark that implements and compares popular

02:34 vector quantizers for the broader community. Quantizers are algorithms that can compress vectors into smaller data representations. Being able to quantize vectors well means you can significantly reduce the cost it takes to store vectors. But, the problem was that it was really hard to implement and compare the performance of different quantizers in the same context and compare [music] them. So, our researchers decided to make a repository that makes it easy to build, evaluate, and measure those quantizers. You can take a look at the website here or the code base here.

03:05 >> [music] >> Finally, the cost export API. We’ve also released a cost and usage export API for [music] enterprise customers. This is handy for grabbing up to a year of Pinecone cost and usage data in the FinOps open cost and usage specification focus format. [music] This is the same format popular clouds they use, and it makes it easy to analyze usage information [music] for finance teams at enterprise organizations in days rather than later. The cost API is available on the Pinecone enterprise plan.

03:35 And that’s a wrap for our change log today. [music] Is there anything you’d like us to bring to Pinecone? Drop a comment below and subscribe so you can get a hold of the next change log. Thanks, everybody. >> [music]