US 12,346,317 B2
Systems and methods for providing a semantic query framework
Anagha Rumade, Mountain View, CA (US); Anjana Umapathy, Santa Clara, NY (US); Sadra Amiri Moghadam, Mission San Jose, CA (US); Abhik Banerjee, Milpitas, CA (US); Gupta Gundlapalli, Frisco, TX (US); Jainesh Doshi, San Francisco, CA (US); and Srinivasa Murthy Basavaraju, San Ramon, CA (US)
Assigned to JPMORGAN CHASE BANK, N.A., New York, NY (US)
Filed by JPMORGAN CHASE BANK, N.A., New York, NY (US)
Filed on Sep. 11, 2023, as Appl. No. 18/464,942.
Prior Publication US 2025/0086170 A1, Mar. 13, 2025
Int. Cl. G06F 16/242 (2019.01); G06F 16/22 (2019.01); G06F 16/334 (2025.01)
CPC G06F 16/243 (2019.01) [G06F 16/2237 (2019.01); G06F 16/3347 (2019.01)] 18 Claims
OG exemplary drawing
 
1. A method comprising:
receiving, from a query interface, a natural language query;
receiving, from an external document source and at a document pool of a document embedding platform, an external document comprising external data;
receiving, from an internal document source and at the document pool, a knowledge graph comprising internal data;
flattening, by a flattening engine of the document embedding platform, the knowledge graph into graph data in a document format and storing the document format in the document pool;
encoding, by a document encoding module of the document embedding platform, the text of the document pool into tokens that represent sub-portions of the text;
embedding, by a document embedding module of the document embedding platform, the tokens into a natural language query vector, wherein the natural language query vector represents a semantic meaning of one or more documents of the document pool;
formatting, by an index query of an index datastore of the document embedding platform, wherein the index query includes the natural language query vector as a lookup parameter of the index query;
storing, as a response to the index query at the index datastore, a document set as a binary large object including a relation to another associated document;
encoding, by a query encoder of a semantic query framework, the natural language query into tokens that represent sub-portions of the natural language query;
generating, by a machine learning model of the semantic query framework, a response to the natural language query based on the document set;
generating, by a fact checking engine of the semantic query framework, a comparison of the response and the document set;
passing, by the document pool or the index datastore, the document set to the machine learning model based on the relation and the index query;
resolve conflicts, by a conflict resolver of the semantic query framework, by selecting internal data for the response where internal data included in the response conflicts with external data in the response;
removing, by the conflict resolver, the external data from the response; and
displaying, by a query interface of the semantic query framework, the response.