This page is the support map for the DSL Query API: for every query type, aggregation, and option, it shows how faithfully each search engine backend (Elasticsearch, Solr, Lucene) implements it. Because Viglet Turing exposes one Elasticsearch-compatible syntax over three different engines, a feature that’s native on one may be translated or only partially supported on another.
Reach for this table when you’re choosing a backend for a site, or when a DSL query behaves differently than you expected across engines: the badge tells you whether that’s a 1:1 mapping, a translation, or a best-effort approximation. Read it alongside the query examples on the DSL Query API page.
Content AI is a federated, read-only index whose query
language is Adobe’s own: a filter DSL plus semantic and keyword queries, not Lucene
syntax. So rather than approximate the whole Elasticsearch surface, Turing translates
a defined subset and refuses the rest by name, with 501 Not Implemented:
Query / feature
Content AI
Notes
match_all
NOT exists(...) over a reserved field name: an empty filter group is invalid in Adobe’s schema
match
A scoring fulltext query
term
StringFilter, NumberFilter or BooleanFilter, chosen from the value’s type
terms
InFilter
range
An AND of one-sided NumberFilter / DateFilter bounds
exists
ExistsFilter
prefix
PrefixFilter
bool
Text clauses become the arms of a composite; the rest becomes one filter group used as their pre-filter. minimum_should_match > 1 and a must_not over a text query are refused
terms aggregation
termBuckets
Sorting
orderBy; _score is dropped, since Adobe already sorts by descending score
size
Capped at 50 by the API
from (offset)
Adobe paginates by opaque cursor; a non-zero from is refused rather than answered with page one
Everything else
knn, nested, geo, function_score, span, join queries, other aggregations
Why refuse rather than approximate. A clause quietly dropped because it could not be
translated would leave a query that still runs, and answers confidently from the wrong
set of documents. A 501 naming the clause is a worse response and a better answer.
This subset is what the vectorless copilot needs, which is the point: that surface
needs no embeddings, so it is the grounded-answer path that suits a federated index, and
it runs entirely through this API.
