Observability for Vector Search: Why Your Green Dashboard Is Lying To You

Dashboards are green – steady QPS, zero 5xx errors, ultra-low latency – yet users report AI hallucinations. Welcome to the silent nightmare of ANN search: recall degradation without infrastructure failure. How do you detect invisible failures? This session maps the four tiers of vector observability: DB internals, index health, LLM performance, and output quality. We’ll explore a production OSS stack (Percona for MongoDB, PMM, mongot, vLLM) and expose where open source still falls short.

Speakers

  • Radosław Szulgo
    Radosław Szulgo
    Percona

    Radoslaw Szulgo is a Senior Product Manager at Percona with more than 15 years of experience in the APM, observability, and distributed systems domain as a software developer, leader, and product manager. Currently, he helps organizations to innovate with MongoDB. Outside of work, he plays in an amateur volleyball league and is passionate about self-development.

  • Sandra Romanchenko
    Sandra Romanchenko
    Percona

    Building on a strong foundation in technical support, Sandra is currently a Staff QA Engineer at Percona, where her analytical problem-solving and drive to dig deep into how systems work shape the automation framework for MongoDB.

Date

Nov 18 2026

Time

17:30 - 18:00

Location

Ballroom 3