AI search

AI search lets you describe what you’re looking for in natural language instead of building a structured query by hand. You type a question; Dataerai turns it into the same precise metadata conditions you’d otherwise assemble in Search & discover, runs it, and returns matching assets.

​

How it works

1. You enter a natural-language request, e.g. “recent SEM or TEM studies”.
2. Dataerai reads the statistics of the data you can see — which fields exist, their common values, their ranges — and a language model designs a structured query that captures each facet of your request.
3. It enumerates real alternatives as OR branches (SEM or TEM), combines facets with and, and excludes with not — every value drawn from your actual data, not invented.
4. Each branch is probed for a match count, and you see the designed query — editable chips, or a read-only OR-tree when it nests — with counts, before it runs against the assets you can already access. AI search never widens your permissions.

Because the translation targets a strict query grammar grounded in real values, the model decomposes rather than guesses. A vague request with no structured facet (“show me interesting papers”) falls back to topical similarity rather than inventing filters and surfacing the wrong assets.

​

Reading the results

When a translation is worth explaining, an insights panel appears beneath the search box before you go to the results, so you see the query that was designed and how it landed:

• Per-condition matches — a count next to each branch of the query. An amber “no matches” tells you a value you expected isn’t in your data before you commit.
• Grounding misses — if a “like this” phrase resolves to no value you can see, you get a badge saying the search was widened to a text search, instead of a silently empty page.
• No exact matches — if a precise query matches nothing, Dataerai widens to the closest results by meaning and shows a “No exact matches — showing the closest results by meaning” banner.

For a flat query you get editable chips you can refine; a query that uses nested OR/NOT logic shows as a read-only tree with a Show results button. The deep dive lives in Grounded OR-tree search.

​

Summarize selected assets

When AI search is enabled, you can also ask Dataerai to summarize a set of assets in plain language. Select up to 50 assets in the results, then click Summarize in the summary panel. The summary streams in as it’s generated — click Stop at any time to keep what’s arrived so far. The panel notes how many assets were included and flags any that were skipped because you don’t have access or were too large to read in full.

​

What makes a good query

​

Topical & visual matching

For “find me things about X” questions, AI search can rank figures by visual similarity rather than keyword overlap — so “biological samples” returns images that look biological, not just ones whose caption happens to share the words. This lane is off by default and is enabled by your administrator. It also requires that your images have been processed for similarity search. Until then, topical queries fall back to the structured translation. A related value-similarity lane lets you describe a metadata value you can’t name precisely (“scanning-electron-type techniques”) and have Dataerai match it to the real values that exist in your data. It is enabled separately by your administrator; see the “like this” examples in Grounded OR-tree search.

​

Availability

AI search appears only when your organization has opted in to AI features. If you don’t see the AI search input, ask an organization admin to enable it — see AI policy & access.

​

Next steps