US 12,412,077 B2
Systems and methods for managing decentralized data sources in generative artificial intelligence pipelines
Peter Mondlock, Highland Park, IL (US); William M.P. Hudelson, London (GB); and Kitti Lakner, New York, NY (US)
Assigned to MCKINSEY & COMPANY, INC., New York, NY (US)
Filed by MCKINSEY & COMPANY, INC., New York, NY (US)
Filed on Oct. 24, 2023, as Appl. No. 18/493,741.
Prior Publication US 2025/0131247 A1, Apr. 24, 2025
Int. Cl. G06N 3/0455 (2023.01); G06F 16/2458 (2019.01); G06F 16/28 (2019.01); G06N 3/0475 (2023.01)
CPC G06N 3/0455 (2023.01) [G06F 16/2471 (2019.01); G06F 16/284 (2019.01); G06N 3/0475 (2023.01)] 11 Claims
OG exemplary drawing
 
1. A computer-implemented method for providing a decentralized knowledge engine for use with a plurality of external third-party data sources to respond to a user query, the computer-implemented method comprising:
receiving, by one or more processors, the user query from a user device;
generating, by the one or more processors, a user query embedding of the user query;
determining, by the one or more processors, an intent associated with the user query by comparing the user query embedding to a plurality of intent embeddings;
selecting, by the one or more processors, one or more relevant external third-party data sources from the plurality of external third-party data sources by comparing (i) an intent embedding representing the intent as a multi-dimensional vector to (ii) a plurality of external third-party data source metadata embeddings representing respective external third-party data source metadata of each of the plurality of external third-party data sources, wherein the one or more relevant external third-party data sources are (i) distinct from the user device and (ii) selected based upon semantic similarity of their respective external third-party data source metadata embeddings to the intent embedding;
retrieving, by the one or more processors, unstructured text from at least one of the one or more relevant external third-party data sources, wherein the unstructured text is distinct from the user query;
retrieving, by the one or more processors, structured data from at least one of the one or more relevant external third-party data sources;
causing, by the one or more processors, the unstructured text to be split into a plurality of text chunks;
causing, by the one or more processors, the structured data to be split into a plurality of data chunks;
for each text chunk in the plurality of text chunks, sending, by the one or more processors, an augmented text chunk comprising (i) the text chunk, (ii) the user query, and (iii) an extraction prompt to a large language model (LLM) to cause the LLM to extract relevant information from the text chunk;
for each data chunk in the plurality of data chunks, sending, by the one or more processors, an augmented data chunk comprising (i) the data chunk, (ii) the user query, and (iii) an extraction prompt to the LLM to cause the LLM to extract relevant information from the data chunk;
sending, by the one or more processors, one or more augmented user queries to the LLM to cause the LLM to generate a response, the one or more augmented user queries collectively comprising (i) the relevant information, (ii) the user query, and (iii) a prompt; and
outputting, by the one or more processors, the response.