| CPC G06F 40/30 (2020.01) [G06F 16/3344 (2019.01); G06T 11/206 (2013.01); G06T 2200/24 (2013.01)] | 24 Claims |
|
1. A method performed by one or more computer systems, each computer system having at least one processor and a memory, the method comprising:
accessing a first network structure comprising a first plurality of nodes;
initiating a query evaluation of a query in a query language, the query comprising a plurality of subexpressions;
recursively evaluating each subexpression of the plurality of subexpressions, wherein recursively evaluating comprises:
for each subexpression that requires semantic evaluation, evaluating semantically, wherein evaluating semantically comprises:
performing a semantic stored procedure, wherein performing the semantic stored procedure comprises:
creating a semantic stored procedure candidate prompt by:
retrieving a first metaprompt designed to generate a semantic task-describing candidate prompt instructing a large language model to emulate a stored procedure;
specializing the retrieved first metaprompt by template filling task-specification template fields in the metaprompt with semantic arguments associated with the subexpression that requires semantic evaluation, resulting in a filled first metaprompt; and
evaluating the filled first metaprompt with a large language model to produce an unfilled semantic stored procedure candidate prompt that includes unfilled template fields for node-associated attribute data and that specifies a particular form of return output;
constructing a plurality of grouping subsets, each grouping subset comprising one or more nodes selected from the first plurality of nodes of the first network structure;
for each grouping subset of nodes in the plurality of grouping subsets:
accessing data associated with each node in the grouping subset;
substituting relevant attribute data into the corresponding template fields of the unfilled semantic stored procedure candidate prompt to yield a filled semantic stored procedure prompt associated with the grouping subset; and
evaluating the filled semantic stored procedure prompt associated with the grouping subset using a large language model to produce an output associated with the grouping subset; and
collecting the produced outputs over the plurality of grouping subsets; and
returning the collected outputs of the semantic stored procedure to the query evaluation; and
evaluating one or more subexpressions as necessary to resolve the query; and
calculating a second network structure that induces connectivity between nodes in response to the evaluation of the query; and
causing a graph visualization of the second network structure to be rendered in a user interface.
|