DIPr Lab to Appear at Sigmod SeQureDB DB 2026

Lakshmi Sahithi’s paper has been accepted to the ACM SIGMOD/PODS Conference (SeQureDB) 2026. SeQureDB 2026 takes place on May 31 - June 05, 2026 in Bengaluru, India.

This paper PAVS - Policy Aware Vector Search, studies how fine-grained access control policies can be enforced in vector databases while preserving both recall, policy correctness with minimum possible latency. We evaluate different enforcement strategies, including pre-filtering, post-filtering, iterative post filtering, and analyze their trade-offs across recall, latency, selectivity, and policy complexity. We propose a new approach Parallel Post Filtering which is proven to achieve higher recall than post-filtering with lesser latency than Iterative post-filtering.

The goal of this project is to make access control as a first-class concern in vector database systems. Instead of applying policies only after retrieval, we explore how vector search and policy enforcement can be co-designed to support secure, efficient, and accurate retrieval for AI applications.

The Key Takeaways:

• We introduce Parallel Filtering, a new enforcement strategy that explores multiple regions of the HNSW index in parallel to improve policy-compliant vector retrieval.
• Our experiments show that pre-filtering provides the highest recall, but it can become expensive when policies are less selective and distances has to computed on large candidate sets.
• Naïve post-filtering is fast but misses authorized results, leading to very low recall when the retrieved nearest neighbors do not satisfy the access-control policy.
• Iterative post filtering performs search in multiple iterations, thus increase in the latency.
• Our Approach Parallel Filtering improves recall over naïve post-filtering while achieving lower latency than iterative post-filtering, making it a middle-ground strategy for post-filtering in fine-grained access control in vector databases.

The full paper is available at - https://doi.org/10.1145/3807894.3810276