Search platforms built before this generation of AI were not designed for the vector ingestion rates, hybrid search throughput, or agent-scale relevance evaluation that production AI applications now demand. Teams running semantic search or retrieval-augmented generation pipelines hit these ceilings fast, and the gap between benchmark performance and real-world throughput is where workloads break.
In this interview on TFiR, Bianca Lewis, Executive Director at the OpenSearch Software Foundation, walks through the concrete performance improvements in the latest OpenSearch release and explains why iterative infrastructure upgrades are the practical path to AI-scale search.
Guest: Bianca Lewis, Executive Director at OpenSearch Software Foundation
Show: Let’s Talk
LinkedIn: Bianca Lewis
Watch the discussion on YouTube
Here is what every platform engineer and search infrastructure team needs to know.
Technical Deep Dive
Q: What does the latest OpenSearch release actually improve for production AI workloads?
Bianca Lewis, Executive Director at the OpenSearch Software Foundation, states that the latest release focuses on real-world improvements and new capabilities rather than narrow benchmark gains. Lewis describes the changes as iterative advances that move search infrastructure further along AI journeys at the data layer, covering vector ingestion speed, search throughput, agent integrations, relevance evaluation, and log analysis tooling.
“It’s not about achieving 0.3-millisecond performance in this use case at midnight; it’s about practically advancing our AI journeys on the data infrastructure.”
Bianca Lewis, Executive Director, OpenSearch Software Foundation
Q: How much faster is vector ingestion in the latest OpenSearch version?
The latest OpenSearch release ingests vectors at 4.16 times the speed of the previous version. Lewis attributes this improvement to changes including base64 encoding of vector payloads. This level of throughput increase directly affects teams running retrieval-augmented generation or semantic search pipelines at scale, where ingestion bottlenecks previously limited index freshness and query latency.
“We ingest vectors now at 4.16 times faster than the previous version.”
Bianca Lewis, Executive Director, OpenSearch Software Foundation
Q: What happened to radial search throughput in this OpenSearch release?
Radial search throughput has more than doubled compared to the previous version of OpenSearch. Lewis positions this alongside the vector ingestion improvement as part of a broader effort to scale the platform’s core search capabilities to meet the demands of modern AI applications, not just incremental feature additions.
“Improving radial search throughput by more than double.”
Bianca Lewis, Executive Director, OpenSearch Software Foundation
Q: How does OpenSearch now integrate with AI agents and model context protocols?
The release extends OpenSearch’s Model Context Protocol (MCP) integrations to support additional agent types beyond what was available previously. This allows search to function as an active component in agent-driven AI workflows rather than a passive data store, giving agent systems structured access to indexed content through standardized protocol interfaces.
“We’ve extended our model context protocol integrations to even more agent types.”
Bianca Lewis, Executive Director, OpenSearch Software Foundation
Q: Is OpenSearch locked to a specific LLM provider for relevance evaluation?
OpenSearch is LLM-agnostic for relevance evaluation. Teams can scale relevance scoring using open-weight models or hosted API providers, with no requirement to commit to a single vendor. Lewis frames this as flexibility to match existing infrastructure choices rather than forcing a provider dependency into the search layer.
“Whether you use open weight or APIs, we’re agnostic to the LLM models that you’re actually using.”
Bianca Lewis, Executive Director, OpenSearch Software Foundation
Q: What has changed for log analysis in this OpenSearch release?
OpenSearch now ships a visual PPL (Piped Processing Language) builder out of the box, covering SQL queries and an onboarding canvas. The tooling includes support for time series data comparison and new PPL commands, making log analysis workflows accessible without requiring users to write raw query syntax. Lewis describes this as a streamlining of the log analysis experience within the platform.
“We’ve streamlined log analysis by using a visual PPL builder which is now in OpenSearch out of the box, which includes SQL queries and an onboarding canvas.”
Bianca Lewis, Executive Director, OpenSearch Software Foundation
Q: Why do search platforms built before generative AI need to modernize now?
Lewis explains that most search and data infrastructure platforms were designed and built before this generation of AI emerged. While AI itself is not new, the current generation of large language models and agent-based applications imposes requirements on vector scale, ingestion speed, and hybrid search throughput that earlier architectures were not built to handle. Iterative modernization is the path to closing that gap without full platform replacement.
“Software platforms were built and designed before the advent of this generation of AI. These iterative changes are really important to achieving the scale needed to deliver real-world applications.”
Bianca Lewis, Executive Director, OpenSearch Software Foundation
Resources & Documentation
- OpenSearch Software Foundation, home of the open-source OpenSearch project, release notes, and documentation
***
👇 Click to Read Full Raw Transcript
Swapnil Bhartiya: We have talked about the community, of course, we talk about the holy co. Let’s talk about the software itself. The latest release was of course it landed right around the conference. Talk a bit of what’s new in this release.
Bianca Lewis: Yeah, Swapna. I think that in the latest release, and the latest releases are about real world improvements which are iterative, about new capabilities. It’s not about performing 0.3 milliseconds in this use case at midnight, it’s about practically getting us further along our AI journeys on the data infrastructure. So some concrete examples of what I’m talking about here is that we’ve extended our model context protocol integrations to even more agent types. More than that, we ingest vectors now at 4.16 times faster than the previous version and improving radial search throughput by more than double. Okay, in terms of. Then we can of course scale the relevance evaluation with the access to LLMs and providers. Whether you use Open Weight or APIs, we’re agnostic to the LLM models that you’re actually using. And then of course we’ve streamlined log analysis by using a visual PPL builder which is now in, in open search out of the box, which includes SQL queries and an onboarding canvas for you to use and things like this, which actually makes it pretty easy to compare things like time series data, new PPL commands and things like this. So the real world application on scaling vector search, smart search, hybrid search, relevance training, log analysis have all scaled way beyond just a couple of versions back. And that modernization is incredibly important because as we know, software platforms were built and designed before the advent of this generation AI. AI is not new, it’s been around for a long time. But this generation of AI, so these iterative changes of invest in vectors four times faster, okay, with base 64 enclosed and things like that, and these are really important to reach the scales that we need to deliver real world applications.





