Granted Patent B1

US 12,547,653 · App. 18/948,835 · Granted Feb 10, 2026

Inventors: Narasimha Sekhar Kakaraparthi (Hyderabad, IN); Rajeev Ranjan Kumar (Hyderabad, IN); Seshu Kumar G V (Hyderabad, IN)

Assignee: Wipro Limited

G06F16/35G06F40/284

Abstract

Claims (86)

1 . A method of identifying domain-specific service management hotspots, the method comprising:

receiving, by a causal analysis server, raw data corresponding to a plurality of domain-specific issues, wherein the raw data comprises at least one of image data and text data;

generating, by the causal analysis server, a plurality of chunks from the raw data using an optimal chunking and tokenizing algorithm selected from a set of chunking and tokenizing algorithms to obtain chunked data;

creating, by the causal analysis server, a plurality of embeddings from the chunked data through an embedding model;

clustering, by the causal analysis server, the plurality of embeddings into a set of clusters and a set of sub-clusters associated with each of the set of clusters, based on a similarity analysis;

determining, by the causal analysis server, a set of cluster labels corresponding to the set of clusters and a set of sub-cluster labels corresponding to the set of sub-clusters associated with each of the set of clusters, via a Large Language Model (LLM), wherein the set of cluster labels and the set of sub-cluster labels are in natural language;

identifying, by the causal analysis server, a plurality of hotspots corresponding to the plurality of domain-specific issues based on the set of cluster labels and the set of sub-cluster labels;

receiving, by a Graphical User Interface (GUI), a user feedback corresponding to one or more of the set of cluster labels or one or more of the set of sub-cluster labels corresponding to each of the set of cluster labels;

creating a prompt based on the user feedback;

inputting the prompt to the LLM; and

for each label of the one or more of the set of cluster labels or one or more of the set of sub-cluster labels corresponding to each of the set of cluster labels, generating, by the LLM, an updated label based on the prompt.

2 . The method of claim 1 , further comprising:

pre-processing the raw data using one or more pre-processing techniques, wherein the pre-processing comprises:

when the raw data comprises the image data, obtaining extracted text data from the image data through a text extraction technique;

performing data cleaning on the text data of the raw data and the extracted text data; and

removing Personally Identifiable Information (PII) data from the text data and the extracted text data.

3 . The method of claim 1 , further comprising:

for each chunking and tokenizing algorithm of the set of chunking and tokenizing algorithms,

generating a preliminary plurality of chunks from the raw data through the chunking and tokenizing algorithm, wherein each of the set of chunking and tokenizing algorithms is configured with a set of parameters;

generating a preliminary plurality of tokens from the preliminary plurality of chunks through the chunking and tokenizing algorithm; and

evaluating the preliminary plurality of tokens based on a cosine similarity score variance; and

selecting the optimal chunking and tokenizing algorithm from the set of chunking and tokenizing algorithms based on evaluation of the preliminary plurality of tokens, wherein the plurality of chunks is the preliminary plurality of chunks generated by the optimal chunking and tokenizing algorithm.

4 . The method of claim 1 , wherein creating the plurality of embeddings from the chunked data comprises:

receiving the plurality of chunks generated by the optimal chunking and tokenizing algorithm;

generating a plurality of tokens from the plurality of chunks through the optimal chunking and tokenizing algorithm;

creating the plurality of embeddings from the plurality of tokens through the embedding model; and

reducing dimensions of the plurality of embeddings through a dimensionality reduction algorithm preconfigured with a set of customized hyperparameters.

5 . The method of claim 1 , wherein the similarity analysis comprises:

calculating a similarity score corresponding to the plurality of embeddings; and

based on the similarity score, grouping one or more of the plurality of embeddings into one of:

a cluster of the set of clusters, or

a sub-cluster of the set of sub-clusters associated with each of the set of clusters.

6 . The method of claim 1 , wherein determining the set of cluster labels and the set of sub-cluster labels comprises:

creating a prompt based on the set of clusters and the set of sub-clusters associated with each of the set of clusters;

inputting the prompt to the LLM, wherein the LLM is a fine-tuned LLM specific to a domain of the plurality of domain-specific issues; and

generating, using the LLM, the set of cluster labels corresponding to the set of clusters and the set of sub-cluster labels corresponding to the set of sub-clusters based on the prompt.

7 . The method of claim 1 , wherein the raw data comprises at least one of ticket dump data, Configuration Management Database (CMDB) data, and logs and events data.

8 . The method of claim 7 , further comprising:

identifying one or more affected servers corresponding to each of the set of sub-clusters using the CMDB data; and

determining a count corresponding to each of the set of sub-clusters using the ticket dump data.

9 . The method of claim 8 , further comprising:

rendering, using the GUI, a visualization map based on the set of cluster labels and the set of sub-cluster labels, the one or more affected servers, and the count.

10 . A causal analysis server for identifying domain-specific service management hotspots, the system comprising:

a processor; and

a memory communicatively coupled to the processor, wherein the memory stores processor instructions, which when executed by the processor, cause the processor to:

receive raw data corresponding to a plurality of domain-specific issues, wherein the raw data comprises at least one of image data and text data;

generate a plurality of chunks from the raw data using an optimal chunking and tokenizing algorithm selected from a set of chunking and tokenizing algorithms to obtain chunked data;

create a plurality of embeddings from the chunked data through an embedding model;

cluster the plurality of embeddings into a set of clusters and a set of sub-clusters associated with each of the set of clusters, based on a similarity analysis;

determine a set of cluster labels corresponding to the set of clusters and a set of sub-cluster labels corresponding to the set of sub-clusters associated with each of the set of clusters, via a Large Language Model (LLM), wherein the set of cluster labels and the set of sub-cluster labels are in natural language;

identify a plurality of hotspots corresponding to the plurality of domain-specific issues based on the set of cluster labels and the set of sub-cluster labels;

receive, by a Graphical User Interface (GUI), a user feedback corresponding to one or more of the set of cluster labels or one or more of the set of sub-cluster labels corresponding to each of the set of cluster labels;

create a prompt based on the user feedback;

input the prompt to the LLM; and

for each label of the one or more of the set of cluster labels or one or more of the set of sub-cluster labels corresponding to each of the set of cluster labels, generate, by the LLM, an updated label based on the prompt.

11 . The causal analysis server of claim 10 , wherein the processor instructions, on execution, further cause the processor to:

pre-process the raw data using one or more pre-processing techniques, wherein for pre-processing, the processor is configured to:

when the raw data comprises the image data, obtain extracted text data from the image data through a text extraction technique;

perform data cleaning on the text data of the raw data and the extracted text data; and

remove Personally Identifiable Information (PII) data from the text data and the extracted text data.

12 . The causal analysis server of claim 10 , wherein the processor instructions, on execution, further cause the processor to:

for each chunking and tokenizing algorithm of the set of chunking and tokenizing algorithms,

generate a preliminary plurality of chunks from the raw data through the chunking and tokenizing algorithm, wherein each of the set of chunking and tokenizing algorithms is configured with a set of parameters;

generate a preliminary plurality of tokens from the preliminary plurality of chunks through the chunking and tokenizing algorithm; and

evaluate the preliminary plurality of tokens based on a cosine similarity score variance; and

select the optimal chunking and tokenizing algorithm from the set of chunking and tokenizing algorithms based on evaluation of the preliminary plurality of tokens, wherein the plurality of chunks is the preliminary plurality of chunks generated by the optimal chunking and tokenizing algorithm.

13 . The causal analysis server of claim 12 , wherein to create the plurality of embeddings from the chunked data, the processor instructions, on execution, further cause the processor to:

receive the plurality of chunks generated by the optimal chunking and tokenizing algorithm;

generate a plurality of tokens from the plurality of chunks through the optimal chunking and tokenizing algorithm;

create the plurality of embeddings from the plurality of tokens through the embedding model; and

reduce dimensions of the plurality of embeddings through a dimensionality reduction algorithm preconfigured with a set of customized hyperparameters.

14 . The causal analysis server of claim 10 , wherein for similarity analysis the processor instructions, on execution, further cause the processor to:

calculate a similarity score corresponding to the plurality of embeddings; and

based on the similarity score, group one or more of the plurality of embeddings into one of:

a cluster of the set of clusters, or

a sub-cluster of the set of sub-clusters associated with each of the set of clusters.

15 . The causal analysis server of claim 10 , wherein to determine the set of cluster labels and the set of sub-cluster labels the processor instructions, on execution, further cause the processor to:

create a prompt based on the set of clusters and the set of sub-clusters associated with each of the set of clusters;

input the prompt to the LLM, wherein the LLM is a fine-tuned LLM specific to a domain of the plurality of domain-specific issues; and

generate, using the LLM, the set of cluster labels corresponding to the set of clusters and the set of sub-cluster labels corresponding to the set of sub-clusters based on the prompt.

16 . The causal analysis server of claim 10 , wherein the raw data comprises at least one of ticket dump data, Configuration Management Database (CMDB) data, and logs and events data.

17 . The causal analysis server of claim 16 , wherein the processor instructions, on execution, further cause the processor to:

identify one or more affected servers corresponding to each of the set of sub-clusters using the CMDB data; and

determine a count corresponding to each of the set of sub-clusters using the ticket dump data.

18 . The causal analysis server of claim 17 , wherein the processor instructions, on execution, further cause the processor to:

render, using the GUI, a visualization map based on the set of cluster labels and the set of sub-cluster labels, the one or more affected servers, and the count.

Assignments (1)

Priority Claims (1)

References Cited (17)