Research

Podcast: EviDex — Chasing the Truth Before It Expires

Hill Research |

In the sixth episode of AI on the Hill, host Vivian speaks with Shuang Cao about EviDex, a retrieval system built so that an answer’s freshness is part of whether it is correct at all. Shuang Cao is co-first author on the EviDex paper, which was presented at ACM SIGMETRICS 2026; she led the project at Hill Research with Rui Li before joining Apple, and appears on the episode in her personal capacity, speaking only for herself. The episode runs 17 minutes.

The Assumption Worth Arguing With

Most retrieval systems are built on the premise that an answer can be correct as of its snapshot, and that staleness is a separate problem to be cleaned up later. The episode’s framing line — “freshness is part of correctness” — rejects that split.

The clinical workload behind the research is where the split stops holding. At a fifteen-minute horizon, 31.2% of time-sensitive clinical queries depend on evidence committed within that window. At that rate a snapshot is not a slightly older answer. It is a different answer.

What EviDex Does Differently

Rather than rebuilding an index on a timer and serving whatever the last refresh captured, EviDex maintains a live map of the evidence itself. A new claim recomputes the part of the map it touches instead of triggering a scheduled global rebuild, so freshness becomes a function of work done rather than of a clock. Every answer comes back with the chain of sources and timestamps that produced it.

The Model Is the Last Step

The point Shuang Cao returns to is where the difficulty actually sits. The language model is the final stage, not the engine: the hard work is in retrieval and indexing, and the model’s job is to turn assembled evidence into a readable answer with citations. It does not decide what counts as current.

Relevance to TriClick: Regulated clinical work asks for two things that are usually built separately — evidence that is current, and an answer a reviewer can trace back to the versioned source that supported it. EviDex treats those as one requirement, which is the same discipline TriClick applies across its clinical workflows: an output nobody can trace back to a source is not finished work.

Two Questions for Any AI Answer

The episode closes on advice aimed at anyone using an AI assistant, not only at people building one: do not confuse confidence with truth. Ask two questions of every answer — where did this come from, and how recently was it checked.

AI on the Hill is a series exploring the people and research behind Hill Research’s work in clinical AI.

Learn More