This page is the deep dive, with a gallery of examples. For the short version and how to turn AI search on, see AI search.
What changed
Earlier, a question like “electron microscopy data” often became a single fuzzy text match on the phrase — a rephrasing of what you typed. That misses the point: your corpus might record electron microscopy astechnique = "SEM", "TEM", or "STEM", and a phrase match finds none of them reliably.
Grounded search instead enumerates the real values:
AFM is in the corpus too, but it isn’t electron microscopy, so it’s left out.
Anatomy of a grounded search
1
Evidence
Dataerai assembles the metadata fields you can see, each with its top values (and how often they occur), numeric/date ranges, and how discriminating the field is. A field that splits your corpus many ways (like
technique) ranks above one that’s the same everywhere.2
Design
The model maps each facet of your question onto those fields and enumerates alternatives as
or branches, conjunctions as and, exclusions as not. It prefers exact values from the evidence over guesses.3
Verify
Each top-level branch is probed against your visible data for a match count. A branch that matches nothing is flagged, and Dataerai makes one corrective pass to repair or widen it.
4
Show
You see the designed query — as editable chips when it’s a simple list, or as a read-only OR-tree when it nests — with a count beside each branch, before anything runs.
Grounding never widens your permissions. Every value, statistic, and result is drawn from assets you can already access — a value another team can see but you can’t will never appear in your query, your counts, or your suggestions.
Example gallery
Each example shows what you type and the query Dataerai designs. The exact values depend on your corpus — these use a materials-science workspace.Enumerate matching values — “electron microscopy data”
Enumerate matching values — “electron microscopy data”
A topical term fans out over the listed values that fit it.Why:
technique has values SEM ×1240, TEM ×890, STEM ×310, AFM ×95. The first three are electron microscopy; AFM isn’t, so it’s excluded.Combine a list with a range — “recent SEM or TEM studies”
Combine a list with a range — “recent SEM or TEM studies”
A value-OR ANDed with a numeric threshold read from the field’s distribution.Why:
year ranges 2018–2026 with a median of 2024, so “recent” anchors to >= 2024.Discriminate by field presence — “stained liver or kidney samples”
Discriminate by field presence — “stained liver or kidney samples”
When a field’s mere presence separates a population, grounded search uses Why: tissue is enumerated from real values; “stained” maps to has a
exists.staining field, not a specific value.Exact values plus a widening arm — “graphene-based samples”
Exact values plus a widening arm — “graphene-based samples”
Listed exact values first, with a Why: both
contains arm to catch close variants.graphene and graphene oxide are real values; the contains arm catches any other graphene-prefixed material the statistics didn’t surface.Exclusion — “everything except retracted work”
Exclusion — “everything except retracted work”
A Why:
not around a real value.status has published ×5000, draft ×900, retracted ×40; the request is the complement of one value.Cross-field conjunction — “Titan or Talos sessions with a HAADF detector”
Cross-field conjunction — “Titan or Talos sessions with a HAADF detector”
Two facets, each grounded, combined with
and.Structured + topical — “bar charts about battery degradation”
Structured + topical — “bar charts about battery degradation”
The structured facet stays structured; the descriptive remainder rides one similarity arm.Why:
chart_type = "bar" is a real value; “battery degradation” has no matching field value, so it becomes a single visual/topical-similarity arm (requires the similarity lane — see AI search).Date window — “data acquired in 2025”
Date window — “data acquired in 2025”
A natural date phrase becomes a half-open range.
Presence and absence together — “datasets without provenance but with a license”
Presence and absence together — “datasets without provenance but with a license”
not_exists and exists in one conjunction.Nested branches — “SEM graphene or TEM silicon studies”
Nested branches — “SEM graphene or TEM silicon studies”
Two complete conditions, each an Why: this nests, so it shows as a read-only OR-tree rather than editable chips — chips can’t yet represent grouped OR/NOT. To narrow it, rephrase.
and, joined by or.Don't know the exact value — “scanning-electron-type techniques”
Don't know the exact value — “scanning-electron-type techniques”
When you describe a value you can’t name precisely, Dataerai resolves your phrase to the real values that exist and that you can see.Why: “scanning electron” is matched against the embedded value space for
technique, and every candidate is checked against your visible data before it’s offered. If nothing visible matches, the request falls back to a text search and you’re told so (see Grounding misses). Requires the value-similarity lane to be enabled.Genuinely descriptive — “fatigue measurement experiments”
Genuinely descriptive — “fatigue measurement experiments”
When no field maps to any facet, a single similarity arm is the right answer.Why: grounded search decomposes when it can and only falls back to pure similarity when the request has no structured facet — the opposite of always rephrasing.
Reading the results panel
When you translate a query, the panel shows you what was designed and how it landed.Per-branch match counts
A count beside each condition. Amber “no matches” means that branch is empty in your data — useful to spot a value that doesn’t exist before you commit.
1000+ and — (unknown) are shown when a count is large or couldn’t be measured in time.Editable chips vs. read-only tree
A flat list of conditions becomes editable chips. A nested OR/NOT query renders as a read-only tree with a note that it can’t be refined as chips yet — rephrase to narrow it.
Grounding misses
If a “like this” phrase resolves to no value you can see, you get an amber badge — “No visible values matched … — widened to a text search” — instead of a silently empty page.
Broadened results
If a precise query matches nothing when you run it, Dataerai widens to the closest results by meaning and shows a “No exact matches — showing the closest results by meaning” banner, rather than a bare empty page.
Built for scale
Grounding stays fast no matter how large the catalog grows. Dataerai profiles the vocabulary of your metadata — the distinct fields and values — not every record, and refreshes those profiles from bounded samples in the background. The query you see, the per-branch probes, and the value matching all cost the same whether your workspace holds thousands of assets or trillions.Availability
Grounded OR-tree design turns on with AI search for your organization. The value-similarity lane (the “like this” matching in the gallery above) is enabled separately by an administrator once value embeddings are ready; until then those phrases fall back to a text search. See AI policy & access and Configuration reference.Next steps
AI search
The overview and how to enable it.
Structured search
The clause-and-operator query grounded search compiles to.
Configuration reference
The flags and knobs admins use to tune the lane.