US 2025/0232134 A1Published application

Conversational language model based content retrieval

AssigneeAmazon Technologies, Inc.

InventorsGirish Shekhar, Zohar Karnin

Life of this patent

Estimating the term…

Loading patent…

Abstract

Devices and techniques are generally described for LM-based content retrieval. First query data including a first request related to first content may be received. First action data associated with the first query data may be determined. First prompt data including a representation of the first query data and data representing the first action data may be generated. The first prompt data may instructs a first LM to recognize entities in the first query data relevant to the first action data. The first LM may determine a first recognized entity from the first request. The first recognized entity may be associated with the first content. A request to resolve the first recognized entity may be generated. A first resolved entity for the first recognized entity may be determined. The first LM may generate first instructions to perform the first action data using the first resolved entity.

Claims20

Dependent claims

A computer-implemented method comprising:

receiving first query data comprising a first natural language request to recommend content of a first type;
performing a lookup using the first query data to determine a first action identifier associated with a first recommendation action;
generating, by a large language model (LLM) orchestrator, first prompt data comprising the first query data and the first recommendation action, wherein the first prompt data instructs a first LLM to recognize entities in the first query data relevant to the first recommendation action;
determining, by the first LLM using the first prompt data, a first recognized entity in the first natural language request, wherein the first recognized entity corresponds to the content of the first type;
sending, by the first LLM to the LLM orchestrator, a request to resolve the first recognized entity;
determining, using a first keyword resolver tool, a first content identifier related to a first instance of the content of the first type and a second content identifier related to a second instance of the content of the first type;
generating, by the LLM orchestrator, second prompt data comprising the first content identifier and the second content identifier;
generating, by the first LLM using the second prompt data, first instructions to perform the first recommendation action using the first content identifier and the second content identifier;
receiving, by an action execution component, the first instructions;
and
executing the first recommendation action by the action execution component, wherein the executing the first recommendation action comprises causes an output of a list comprising data representing the first instance of the content of the first type and data representing the second instance of the content of the first type.

8 more elements

A system comprising:

at least one processor;
and
non-transitory computer-readable memory storing instructions that, when executed by the at least one processor, are effective to:
receive first query data comprising a first request related to first content;
determine, using the first query data, first action data associated with the first query data;
generate first prompt data comprising a representation of the first query data and data representing the first action data, wherein the first prompt data instructs a first language model (LM) to recognize entities in the first query data relevant to the first action data;
determine, by the first LM based at least in part on the first prompt data, a first recognized entity from the first request, wherein the first recognized entity is associated with the first content;
generate a request to resolve the first recognized entity;
determine a first resolved entity for the first recognized entity;
generate second prompt data comprising the first resolved entity;
generate, by the first LM based at least in part on the second prompt data, first instructions to perform at least one action associated with the first action data using the first resolved entity;
and
generate output data associated with the first resolved entity based at least in part on the first instructions.

10 more elements

Dependencies are inferred from the claim text.