Granted Patent B2

US 12,730,963 · App. 18/926,166 · Granted Sep 8, 2026

Inventors: Matthew Kaye (Cambridge, MA); Alex Riina (Cambridge, MA); Manuel Ventero Peña (Cambridge, MA)

Assignee: Klaviyo, Inc.

G06F40/166G06F40/279

Abstract

Claims (55)

1 . A computer-implemented method for automatically providing dynamic content support to a user comprising:

indexing a plurality of articles, wherein the indexing comprises preparing a plurality of pre-existing embeddings corresponding to the plurality of articles or sections thereof using at least one embedding model, and saving the plurality of pre-existing embeddings in a vector database;

receiving a natural language-based input at a user device from the user;

pre-processing the natural language-based input by evaluating the natural language-based input for pre-defined stop words and removing the pre-defined stop words from the natural language-based input if detected;

transform the pre-processed input into at least one current embedding using the at least one embedding model;

compute a similarity score between the at least one current embedding and the pre-existing embeddings corresponding to the articles or sections thereof in the vector database;

ranking the pre-existing embeddings based on similarity scores;

adjusting the ranking of the pre-existing embeddings based on the natural language-based input from the user device;

retrieving a selected plurality of articles or sections thereof based on the adjusted ranking of the articles;

creating a prompt for sending to a text generation model, wherein the creating comprises applying at least one static system-based rule and at least one dynamic user-based rule;

sending to the text generation model the prompt and the selected plurality of articles or sections thereof, and requesting from the text generation model a summary of the selected plurality of articles or sections thereof based on the prompt and the selected plurality of articles or sections thereof; and

sending the summary to the user.

2 . The method of claim 1 , wherein the at least one current embedding comprises a plurality of embeddings, and wherein the method further comprises computing an initial score for each current embedding for each pre-existing embedding.

3 . The method of claim 2 , further comprising, for each pre-existing embedding, averaging the initial scores to obtain the similarity score for the pre-existing embedding.

4 . The method of claim 1 , wherein the indexing comprises chunking the articles or sections thereof into at least one size of chunk.

5 . The method of claim 1 , wherein the pre-processing further comprises evaluating the input for generic words and mapping the generic words into platform words.

6 . The method of claim 5 , wherein the pre-processing comprises determining whether the input is a question based on evaluating the question for symbols and keywords.

7 . The method of claim 6 , wherein the pre-processing further comprises evaluating whether input is an answerable question based on whether the question contains general inquiry information and customer specific information.

8 . The method of claim 1 , further comprising performing one or more of the steps of transforming, computing, selecting, creating, and requesting for different users in parallel using web server replicas.

9 . The method of claim 1 , wherein the at least one static system-based rule is selected from the group comprising: (a) to require the text generation model to limit responses to content contained in the articles; (b) to follow good user experience writing principles to make answers digestible and empathetic to average person; and (c) to require the response to be a valid Markdown such that the response can be presented in a nicely-formatted way.

10 . The method of claim 1 , wherein the dynamic user-based rule comprises evaluating the profile characteristics of the user, and adjusting the prompt based on the user profile characteristics.

11 . The method of claim 1 , wherein the adjusting the ranking of the selected articles or sections thereof is based on literal and contextual information contained in the natural language-based input.

12 . The method of claim 11 , further comprising adjusting the ranking of the selected articles or sections thereof based on detected behavior.

13 . The method of claim 12 , wherein the detected behavior corresponds to a flow-triggered event or electronic input sending success.

14 . The method of claim 1 , wherein the creating the prompt is based on feedback, and wherein the feedback is collected from the user's review of the summary.

15 . The method of claim 1 , further comprising detecting behavior characteristics of the user in real time, and presenting the user candidate suggestions for the pre-processed input based on the behavior characteristics of the user.

16 . The method of claim 15 , further comprising building a user question support database based on recording the user behavior and pre-processed input for each user, and presenting the user the candidate suggestions from the database based on the user behavior and information associated with the recorded user behavior.

17 . The method of claim 4 , wherein each chunk corresponds to one of the pre-existing embeddings, and performing the indexing periodically and not less than daily.

18 . A system for automatically providing dynamic content support to a user comprising:

a trained transformer model for transforming text into embeddings;

a vector database comprising articles, and pre-existing embeddings generated by the transformer model corresponding to the articles or sections thereof, and operable to compute a similarity score between a current embedding and each of the pre-existing embeddings;

at least one network or web server programmed and operable to:

receive natural language user input;

pre-process the user input into a pre-processed input, wherein pre-processing comprises removing predefined stop words;

generate, using the transformer model, at least one current embedding from the pre-processed input;

send the at least one current embedding to the vector database to identify a plurality of selected articles or sections thereof based on similarity scores;

rank the plurality of selected articles or sections thereof based on the similarity scores;

reorder the ranked plurality of selected articles or sections thereof based on the user input;

retrieve the reordered plurality of selected articles or sections thereof;

prepare a prompt command for a text generation model based on applying at least one static system-based rule and at least one dynamic customer-based rule;

prompt the text generation model for a summary by sending to the text generation model the reordered plurality of selected articles or sections thereof and the prompt command; and

send the summary to the user.

19 . The system of claim 18 , further comprising a load balancer module programmed and operable to distribute the pre-processed input amongst the at least one network or web server.

20 . A non-transitory computer-readable medium storing instructions that, when executed by one or more computer processors of a computing platform, cause the computing platform to perform the operations comprising:

index a plurality of articles to form pre-existing embeddings corresponding to the articles or sections thereof;

receive natural language user input;

pre-process the user input into a pre-processed input, wherein the pre-processing comprises removing stop words;

transform the pre-processed input into at least one current embedding;

send the at least one current embedding to the vector database to identify a plurality of selected articles or sections thereof based on similarity scores;

rank the plurality of selected articles or sections thereof based on the similarity scores;

reorder the ranked plurality of selected articles or sections thereof based on the user input;

retrieve the reordered plurality of selected articles or sections thereof;

prepare a prompt command for a text generation model based on applying at least one static system-based rule and at least one dynamic customer-based rule;

prompt the text generation model for a summary of the selected articles or sections thereof by sending to the text generation model the reordered plurality of selected articles or sections thereof and the prompt command; and

send the summary to the user.

Assignments (2)

Continuity (1)

References Cited (27)