RelevanceFeedbackQuery is the eleventh variant of Qdrant's Query type, it is in the OpenAPI document
this repository vendors, and QueryMapping says so in a comment: it has no model, it is newer than the
client's query surface, and reaching it would mean adding it to the REST engine too.
That comment is a milestone waiting to be written, and Qdrant's own 2026 roadmap says it will not stay a
curiosity: relevance feedback is one of the three things they name under agent-native retrieval.
The feature is the loop an agent actually runs. A first search returns candidates, something downstream
judges some of them useful and some not, and the next search is supposed to use that judgement. Today a
caller can express it as recommend with positive and negative examples, which is close and not the
same: recommend takes vectors or point ids and treats them as a target, where relevance feedback takes a
graded response to a query that already happened. A client that offers only the approximation makes the
agent's second call worse than the server can answer.
The work is one variant beside the ten QueryInterface already has, its JSON shape, its protobuf
message, and a DSL entry that reads like the others. The size of it is small; the reason it is a
milestone rather than a line is that the query surface is the part of this client people compare
against the official one, and a variant missing from it is visible.
Exit criterion: a relevance-feedback query built from the DSL validates against Qdrant's OpenAPI
document in the contract tests and returns a different ranking from the same query without feedback,
asserted against a real Qdrant over both engines.
RelevanceFeedbackQueryis the eleventh variant of Qdrant'sQuerytype, it is in the OpenAPI documentthis repository vendors, and
QueryMappingsays so in a comment: it has no model, it is newer than theclient's query surface, and reaching it would mean adding it to the REST engine too.
That comment is a milestone waiting to be written, and Qdrant's own 2026 roadmap says it will not stay a
curiosity: relevance feedback is one of the three things they name under agent-native retrieval.
The feature is the loop an agent actually runs. A first search returns candidates, something downstream
judges some of them useful and some not, and the next search is supposed to use that judgement. Today a
caller can express it as
recommendwith positive and negative examples, which is close and not thesame: recommend takes vectors or point ids and treats them as a target, where relevance feedback takes a
graded response to a query that already happened. A client that offers only the approximation makes the
agent's second call worse than the server can answer.
The work is one variant beside the ten
QueryInterfacealready has, its JSON shape, its protobufmessage, and a DSL entry that reads like the others. The size of it is small; the reason it is a
milestone rather than a line is that the query surface is the part of this client people compare
against the official one, and a variant missing from it is visible.
Exit criterion: a relevance-feedback query built from the DSL validates against Qdrant's OpenAPI
document in the contract tests and returns a different ranking from the same query without feedback,
asserted against a real Qdrant over both engines.