Matt McDonough on Couchbase Hyperscale Vector Index

Q1. Couchbase released some benchmark results. The benchmark shows Couchbase achieving 703 QPS with 369ms latency at 93% recall accuracy, while MongoDB delivered only 2 QPS with 40+ seconds latency. What specific architectural decisions in your Hyperscale Vector Index enable this 350X performance advantage at billion-vector scale?

Couchbase Hyperscale Vector Index uses components from the field of Approximate Nearest Neighbor Search (ANNS). It combines the DiskANN search algorithm with the Vamana directed graph construction algorithm to provide flexibility to run in-memory for smaller datasets or operate across partitioned disks for distributed processing and scaling. Based on the results of the benchmark report, we surmise MongoDB’s significant drop in performance happens because its index for 1 billion data points does not fit into memory. Whereas Couchbase is architected so that only a routing graph stays in memory and the vector partitions sit efficiently on disk, allowing for significantly improved performance at scale. 

Q2. You mention that ‘every second of delay translates to frustrated users, abandoned transactions and lost revenue.’ Can you share examples of how enterprises are using Couchbase’s vector search capabilities in production AI applications, and what business outcomes they’re achieving with sub-second latency?

While it’s still very early days for vector search and AI agentic apps, we have seen great early adoption. Employee-generated content platform Seenit uses Couchbase vector search to let users search and recommend video content from over 500,000 videos stored in Couchbase based on meaning, context and keywords. As a result, Seenit is able to achieve sub-second retrieval times to maintain engagement and reduce content discovery friction. IT data consulting and service provider Jinmu chose Couchbase over MongoDB for its AI Assistant project due to Couchbase’s vector search capabilities, SQL syntax support, and ability to store and process time series data. This allows Jinmu to capture conversations in meetings, write summaries and retrieve semantically relevant information from past meetings. With Couchbase, it is able to achieve 8,000 operations per second with 10ms response times.

Q3. The benchmark tested at billion-vector scale with identical AWS infrastructure for both platforms. As organizations scale their AI applications beyond 1 billion vectors, how does Couchbase’s architecture maintain performance, and what are the total cost of ownership implications compared to alternatives?

Couchbase Hyperscale Vector Index is specifically designed to address large datasets. One way customers can measure TCO is vector-queries-per-dollar. At 1 billion scale, Couchbase can deliver 500,000+ vector queries per dollar, while MongoDB delivers just under 300, a 2000x value gap. And with granular support for vector indexing; developers can achieve better performance, accuracy, memory usage and cost savings through flexible multi-dimensional vector indexing.

It’s also important to note that Couchbase is a multi-purpose data platform. Vector is an important part, and Couchbase can handle a variety of data access methods: SQL++, key/value, full text search, mobile sync, caching and more. This reduces the overall footprint of an enterprise’s data architecture: reducing risk, complexity, duplication, licenses, patches, upgrades, SDKs, etc., all contributing to a lower TCO. 

Q4. Traditional vector databases force organizations to choose between speed and accuracy. Your results show Couchbase delivering both 350X faster throughput AND higher recall accuracy (93% vs 89%) than MongoDB. How does the DiskANN algorithm with Vamana graph construction eliminate this traditional trade-off?

Speed vs. accuracy is always a dial, not something that any vector search can eliminate. Couchbase’s hybrid IVF and Vamana approach simply makes the dial more efficient. Couchbase customers get higher recall at the same latency level compared to the Hierarchical Navigable Small World (HNSW) algorithm. 

MongoDB can raise recall, but doing so requires more RAM and more search nodes, which drives up cost. Couchbase maintains performance at high recall without scaling memory linearly.

The real difference is TCO at scale: Couchbase delivers high recall and high throughput at billion-vector scale without needing to oversize hardware, making it economically viable in production.

Q5. You state that ‘performance bottlenecks in existing vector databases are stalling production GenAI deployments.’ With Couchbase 8.0 now generally available, what are the most common barriers enterprises face when operationalizing GenAI applications, and how do the three distinct vector indexing types in Couchbase 8.0 address these challenges across cloud, edge, and on-premises environments?

Many GenAI pilots stall because teams end up stitching together separate systems for operational data, vector search and full-text search, which increases latency and complexity. Enterprises are discovering how much each factor costs when going to production. As a result, expensive-to-scale vector search is a primary barrier.

Couchbase’s three vector types address different use cases and dataset sizes. Hyperscale is for large scale pure vector search; Composite is for retrieving specific vectors for targeted use cases; and Search combines vector search with multiple data types, such as text and geospatial data, into one hybrid search. Because Couchbase runs in the cloud (Capella), on-prem and at the edge, the same AI workflows work everywhere.

Qx. Anything else you would like to add?

AI is operational now, not just experimental. That means vector search needs to meet the same reliability and performance expectations as any core application feature. Couchbase 8.0 gives teams a single platform for operational data, including SQL++, search and vector retrieval, which means less architecture sprawl and lower total cost of ownership. At Couchbase, we’re focused on helping organizations build AI features that remain fast at scale, whether they are running in the cloud, on-prem or at the edge, not just in a demo environment.

…………………………………………..

Matt McDonough, Couchbase SVP of Product.

Matt is responsible for Couchbase partner strategy and worldwide partner execution including channels, technology partners, solution partners, systems integrators, and cloud service providers.