US 12,339,845 B1
Method and system for task planning and execution based on user query
Dagnachew Birru, Marlborough, MA (US); Ankit Prakash, Mumbai (IN); Ayush Dudhe, Mumbai (IN); Sriram Natarajan, Mumbai (IN); Ashwin Sanjeevan, Mumbai (IN); Muneeswaran I, Mumbai (IN); Saisubramaniam Gopalakrishnan, Mumbai (IN); Harpreet Singh, Mumbai (IN); Hamza Moiyadi, Mumbai (IN); Kanishk Mehta, Toronto (CA); and Vishal Vaddina, Toronto (CA)
Assigned to QUANTIPHI, INC, Marlborough, MA (US)
Filed by Quantiphi, Inc, Marlborough, MA (US)
Filed on Mar. 28, 2024, as Appl. No. 18/621,059.
Int. Cl. G06F 16/2453 (2019.01); G06F 16/2455 (2019.01); G06F 16/2457 (2019.01)
CPC G06F 16/24542 (2019.01) [G06F 16/24564 (2019.01); G06F 16/24578 (2019.01)] 10 Claims
 
1. A method for task planning and execution based on a user query (200), the method comprising:
receiving the user query indicative of a task (202) associated with retrieval of data (204) from a database (206) comprising structured and unstructured data sources (208) and generating a relevant response (210) for the user query;
parsing the user query for determining the task associated with the retrieval of the data indicated in the user query, using a Large Language Model (LLM) (214, 304);
generating a task plan (220) comprising a plurality of sub-tasks (222) for the execution of the task associated with the retrieval of the data, using the LLM, wherein executing the plurality of sub-tasks comprises:
determining an embedding (224) for the task associated with the retrieval of the data,
performing a retrieval task (226) for identifying relevant structured and unstructured data sources (228) from amongst the database for the retrieval of the data, based on the determined embedding,
generating relevance scores (230) for the identified relevant structured and unstructured data sources, based on the determined embedding, and
when the relevance scores are indicative of an unstructured pipeline (232):
triggering an unstructured data query engine (234, 306) for searching the identified relevant unstructured data sources and retrieving the data, based on the determined embedding, and
generating the relevant response based on the retrieved data, using the LLM, or
when the relevance scores are indicative of a structured pipeline (236):
triggering a structured data query engine (238, 308) for searching the identified relevant structured data sources and retrieving the data, based on the determined embedding, and
generating the relevant response based on the retrieved data, or
when the relevance scores are indicative of a hybrid pipeline (240):
triggering a structured data query engine for searching the identified relevant structured data sources and retrieving a first portion of the data (242), based on the determined embedding,
receiving the first portion of the data and the user query at a query transformer (244, 310) for transforming the received user query and the first portion of data,
triggering an unstructured data query engine for searching the identified relevant unstructured data sources and retrieving a second portion of the data (246), based on the determined embedding, and
combining the first portion of the data and the second portion of the data for generating the relevant response thereof.