EDB Postgres® AI Outperforms Vector Databases, Lakehouses, and Document Stores for Agentic AI on Speed, Accuracy, and Cost

EDB Postgres AI keeps intelligence next to the data on one sovereign foundation. No ETL, no copies, no separate vector store. Independent benchmarks show it is faster, more accurate, and lower cost than competing lakehouse, vector, and document-store platforms.

WILMINGTON, Del., July 29, 2026 /PRNewswire/ -- EnterpriseDB (EDB), the leading sovereign data and AI company, today announced that EDB Postgres AI (EDB PG AI) outperforms competing AI data platforms on the measures that decide whether enterprise AI works in production: query latency, accuracy, cost, and data freshness. Independent benchmark testing by renowned benchmarking firm McKnight Consulting Group, commissioned by EDB, compared EDB PG AI against specialized vector databases, lakehouses, document stores, and other managed Postgres platforms, all on normalized enterprise hardware. EDB PG AI led on every test.

IDC projects agentic AI will exceed a quarter of worldwide IT spending by 2029, and Gartner expects a third of enterprise applications to embed agents by 2028, up from under 1% two years ago. That pace breaks the architecture still used by most enterprises, where AI runs on a separate copy of the data. Enterprises are answering by standardizing on Postgres as the foundation for AI-ready data—but how you run Postgres determines whether that agentic foundation actually holds up.

"Across every workload we tested—raw vector search, hybrid queries combining vectors with structured filters, and full end-to-end agent retrieval—EDB Postgres AI led the field, and did so while spending fewer tokens and dollars to get there," said William McKnight, president of McKnight Consulting Group. "What stood out was the consistency. Platforms that separate storage from search degraded sharply at scale. The Postgres-based approach held its performance, and EDB's implementation was the strongest we evaluated."

The new benchmarks reinforce the case EDB made in its June launch: Intelligence belongs next to the data, not in a copy held in a lakehouse or a separate vector store.

Millisecond speed at scale, so agents act in the moment

EDB PG AI delivered the fastest query performance of any vendor that McKnight evaluated. While lakehouse and document-store platforms degraded into multi-second response times at scale, EDB PG AI held steady in milliseconds. For a fraud-scoring agent, 50 milliseconds versus three seconds is the difference between blocking a fraudulent transaction and paying for it after it clears.

• 50 ms median query latency at 50 million vectors, the fastest of any platform tested
• 80x faster than Databricks and 21x faster than MongoDB Atlas
• Up to 2x faster than the nearest managed Postgres alternatives, including Aurora and Crunchy Bridge

More accurate results, so every agent acts on the right data

EDB PG AI returned the most accurate results in core vector search of any platform tested, at every scale. Accuracy is not an abstraction for an autonomous agent: The data it retrieves is the data it acts on. A less accurate retrieval layer doesn't just produce a slightly worse answer; it produces a confident wrong action, at machine speed, with no human in the loop to catch it.

EDB PG AI delivers:

• 0.911 Recall@10, the highest accuracy of any platform for core vector search
• 26% higher accuracy than MongoDB and 17% higher than Databricks
• Highest recall of any Postgres platform tested, at both 10M and 50M scale
• More accuracy than open source PostgreSQL itself, providing proof that how you run Postgres decides the outcome

End-to-end agentic retrieval, in one system

To test how enterprise agents fire multiple retrievals and act on the combined results, McKnight modeled a concurrent three-arm agent loop: filtered vector search over patient vitals, filtered vector search over clinical notes, and full-text search over trial protocols run as a single workload.

EDB PG AI completed the full loop in 27 milliseconds, while platforms that split storage from search stretched into hundreds of milliseconds or seconds and lost accuracy exactly when vectors and filters combined (MongoDB's recall fell to 0.21). Because every arm runs in one database on live data, there's no fanning out across systems and no synced copy to fall behind.

• 27ms median agent-loop latency (40ms p99) versus 699ms for MongoDB Atlas and 5.1 seconds for Databricks
• Vector, filtered, and full-text retrieval in a single query path, on one live dataset, with no separate vector store to sync
• 99.7% faster write-to-read than Databricks (12 ms), so the loop always runs on current data
• 67% fewer wasted tokens than Databricks and 58% fewer than MongoDB, so the loop retrieves the right context, not just more of it

Better price performance

A lower hourly rate doesn't mean a lower bill. A slower platform burns more compute and more tokens to do the same work. Once you measure what it actually costs to get the job done, EDB PG AI was the most economical platform tested.

• 76x better price performance than Databricks and 34x better than MongoDB
• 23%–28% better price performance than the nearest managed Postgres competitors, once query speed is accounted for

One sovereign foundation for transactions, analytics, and AI

"These results come from a single architectural choice: We run vectors, structured data, and analytics together, in the operational database, on the live data," said Max Romanenko, chief engineering officer, EDB. "Stand up a separate vector store and you've added a system to secure and pay for, a copy of your data that drifts out of sync and outside your governance, and one more thing that slows down as load climbs. We don't pay that tax, because the intelligence and the data are in the same place. That isn't something you bolt on later. It's structural."

EDB Postgres AI runs operational, analytical, and agentic workloads on one open Postgres foundation, governed at the data layer. This is the architecture Forrester recognized in naming EDB a Leader in The Forrester Wave™: Multimodel Data Platforms, Q2 2026. These benchmarks show that how you run Postgres is what turns the industry's standard into a decisive advantage.

To read the full benchmark report, visit: https://www.enterprisedb.com/agentic-vector-database-benchmark.

About the benchmark

The study was conducted by McKnight Consulting Group and commissioned by EDB. Testing measured pgvector performance across EDB Postgres AI, Amazon Aurora PostgreSQL, Databricks, Crunchy Bridge (Snowflake Postgres), MongoDB Atlas, and open source PostgreSQL, with every platform normalized to equivalent enterprise-grade hardware, across three workloads—core vector indexing, hybrid retrieval and filtered search, and end-to-end agentic retrieval. The full report is available at https://www.enterprisedb.com/agentic-vector-database-benchmark.

About EDB

EDB Postgres AI is the sovereign data and AI platform for the agentic enterprise. Built on Postgres, the world's leading open source database, EDB Postgres AI unifies transactional, analytical, and AI workloads in a single architecture—eliminating the data movement, ETL, and operational fragmentation that slow enterprises down. With governance enforced at the data layer and the flexibility to deploy on-premises, in hybrid environments, or across clouds, enterprises operationalize their data and AI on infrastructure they own and control—reaching production-ready sovereign AI in weeks, not months. As one of the most active contributors to the PostgreSQL project, EDB is deeply invested in the vitality of the global open source community. To learn more, visit www.enterprisedb.com.

EnterpriseDB and EDB are registered trademarks of EnterpriseDB Corporation. Postgres and PostgreSQL are registered trademarks of the PostgreSQL Community Association of Canada and used with their permission. All other trademarks are owned by their respective owners. 

Media Contact:
Anwar Ahmad-Hamdan
[email protected]
(347) 206-5651

SOURCE EnterpriseDB