Question answer (QA) pair generation using agentic workflow system and method that aligns artificial intelligence model with domain-specific principles
A question and answer (QA) pairs generation system and method uses an agentic workflow system and method to align generative artificial intelligence (a large language model (LLM) or a large multimodal model (LMM)) with the principles of a specific domain so that the generative artificial intelligence is better able to respond to a user query in the specific domain.
1 . A method, comprising:
retrieving, by a computer system, a plurality of pieces of content that together embody one or more principles for a specific domain;
creating, on the computer system for each piece of retrieved content, a structured representation of one or more pieces of data in each piece of retrieved content;
performing, on the computer system on the structured representation for each piece of retrieved content, a recursive summarization to generate a summary for each piece of retrieved content that is stored back into the structured representation for each piece of retrieved content;
performing, by the computer system based on the structured representation for each piece of retrieved content that includes the summary, a relevance assessment of each node in the structured representation of the one or more pieces of data in each piece of retrieved content to identify a node relevant to a fine tuning direction for an artificial intelligence model (AIM) to generate an initial plurality of question answer pairs (QA pairs) aligned to the specific domain;
performing, by the computer system on each initial QA pair, an answer refinement for the relevant node to generate a refined QA pair and an acceptable plurality of QA pairs aligned to the one or more principles of the specific domain; and
adding, by the computer system, the refined QA pair to the summary of the relevant node.
2 . The method of claim 1 , wherein the structured representation of each piece of retrieved content is a document tree.
3 . The method of claim 2 , wherein each node of the document tree stores one of a piece of text and a structural element of the piece of retrieved content.
4 . The method of claim 1 further comprising training the artificial intelligence model (AIM) using the acceptable plurality of QA pairs to align the AIM with the specific domain.
5 . The method of claim 4 further comprising receiving a query and generating an aligned response by the aligned AIM to the received query.
6 . The method of claim 4 , wherein the AIM is one of a large language model and a large multimodal model.
7 . The method of claim 6 , wherein the specific domain is one of an industry standard, a civility score, an enterprise domain, a set of pieces of content from a computer and a blog post.
8 . The method of claim 1 further comprising generating one or more alignment processes using the acceptable QA pairs.
9 . The method of claim 8 further comprising outputting a response to a query from a trained artificial intelligence model (AIM) and adjusting, using the one or more alignment processes, the output response from the trained AIM to generate an aligned response to the query that is aligned to the specific domain.
10 . The method of claim 9 , wherein the trained AIM is one of a large language model and a large multimodal model.
11 . The method of claim 10 , wherein the specific domain is one of an industry standard, a civility score, an enterprise domain, a set of pieces of content from a computer and a blog post.
12 . A system, comprising:
a computer system having a hardware processor and a plurality of lines of instructions executed by the hardware processor so that the hardware processor is configured to:
retrieve a plurality of pieces of content that together embody one or more principles for a specific domain;
create, for each piece of retrieved content, a structured representation of one or more pieces of data in each piece of retrieved content;
perform, on the structured representation for each piece of retrieved content, a recursive summarization to generate a summary for each piece of retrieved content that is stored back into the structured representation for each piece of retrieved content;
perform, based on the structured representation for each piece of retrieved content that includes the summary, a relevance assessment of each node in the structured representation of the one or more pieces of data in each piece of retrieved content to identify a node relevant to a fine tuning direction for an artificial intelligence model (AIM) to generate an initial first plurality of question answer pairs (QA pairs) aligned to the specific domain;
perform, on each initial QA pair, an answer refinement for the relevant node to generate a refined QA pair and an acceptable plurality of QA pairs aligned to the specific domain; and
add the refined QA pair to the summary of the relevant node.
13 . The system of claim 12 , wherein the structured representation of each piece of retrieved content is a document tree.
14 . The system of claim 13 , wherein each node in the document tree stores one of a piece of text and a structural element of the piece of retrieved content.
15 . The system of claim 12 , wherein the hardware processor is further configured to train the artificial intelligence model (AIM) using the acceptable plurality of QA pairs to align the AIM with the specific domain.
16 . The system of claim 15 , wherein the hardware processor is further configured to receive a query and generate an aligned response by the aligned AIM to the received query.
17 . The system of claim 15 , wherein the AIM is one of a large language model and a large multimodal model.
18 . The system of claim 17 , wherein the specific domain is one of an industry standard, a civility score, an enterprise domain, a set of pieces of content from a computer and a blog post.
19 . The system of claim 12 , wherein the hardware processor is further configured to generate one or more alignment processes using the acceptable QA pairs.
20 . The system of claim 19 , wherein the hardware processor is further configured to output a response to a query from a trained artificial intelligence model (AIM) and adjust, using the one or more alignment processes, the output response from the trained AIM to generate an aligned response to the query that is aligned to the specific domain.
21 . The system of claim 20 , wherein the trained AIM is one of a large language model and a large multimodal model.
22 . The system of claim 21 , wherein the specific domain is one of an industry standard, a civility score, an enterprise domain, a set of pieces of content from a computer and a blog post.