Superlinked CEO Daniel Svonava and industry legend Hamel Husain sat down for a lively “Lightning Lesson” titled “Optimize Structured Data Retrieval With Evals,” drawing an online crowd of more than seven hundred engineers and product folks. The duo argued that search evaluation is stuck in the past, because most benchmark queries still look like “hotel nyc” while real users type “top family-friendly hotel near Midtown under 200.”

‍

TL;DR:

Key Topics Covered:

• Why current RAG approaches focus too heavily on unstructured data
• The limitations of text-to-SQL for complex queries with multiple signals
• How to encode metadata (timestamps, coordinates, behavioral data) into embeddings
• Why re-ranking is "mega overused" and often unnecessary
• Building metadata-aware embeddings that understand more than just text
• Case studies from jobs marketplace and fashion e-commerce
• Evaluation challenges for metadata-rich search systems

‍

Main Argument: Instead of relying heavily on filters, boosting, and re-ranking, businesses should encode as much signal as possible directly into their vector embeddings for more efficient and accurate retrieval.

‍

Controversial Take: Re-ranking is often just a "hack" to compensate for poor initial retrieval. In an ideal system, the top 10 results from your database should already be the best 10 results.

‍

Daniel opened with the core problem: modern search stacks mix keyword signals, vector similarity, boosts, filters and rerankers, yet teams often grade each component in isolation. That piecemeal testing hides the way structured fields such as price, category or distance collide with free-text relevance. The lesson promised three immediate skills for closing that gap: spotting blind spots in current evals, understanding how hybrid retrieval interacts with structured fields, and building datasets that reflect messy production traffic.

‍

Hamel took the baton to show what “good” looks like in practice. He suggested scraping real catalogues, augmenting them with synthetic long-tail queries, then letting an LLM judge the first pass so humans only review the tricky false positives. Instrumentation is key; track per-field errors so you can tell when an overly aggressive price filter slashes recall. Hamel also cautioned that automatic graders still miss nuance, so teams should hand-check borderline failures before shipping.

‍

The chat ended with both speakers urging viewers to share their evaluation playbooks. “Everyone’s catalogue is weird in its own way,” Daniel laughed, “so we only learn when we expose the failures.” By the session’s close the Q&A feed was buzzing with requests for the slide deck and sample notebooks, a sign that structured-data evals are finally getting the attention they deserve.

‍

Still hungry for more and need help improving your search? Why not jump on a call with one of our co-founders? Supercharge your search and recommendations with Superlinked!