AI Support Bot Trained on Source Code Capabilities
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing AI models struggle to effectively train on software application source code, often focusing on programming language structure rather than the application's capabilities, making it difficult to generate reliable and authoritative natural-language responses for customer support.
Innovation Solution
The proposed solution involves training an AI model on a dataset comprising the source code of a software application, specifically using an AI-supporting interface that allows the model to detect and utilize the capabilities implemented in the code, enabling it to provide accurate and reliable responses to user queries.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If AI model is trained on structured and unstructured data extracted from text documents, audio, video, databases, then customer support responses can be generated, but the responses are not reliable and accurate because these data sources merely represent what someone has said about the software application
Solution Approach 1:
The patent creates a synthetic training dataset by copying and transforming source code into a format suitable for AI training. The source code is processed to extract function calls, parameters, and return values, then formatted as training examples that mimic real software behavior without requiring direct access to the original codebase during inference.
Solution Approach 2:
The patent introduces an intermediary layer between the source code and the AI model training process. This intermediary consists of a code processor that translates source code into a standardized training format, and a training data generator that creates training examples from the processed code, enabling reliable training without direct code-model interaction.
2Measurement precision
If AI model is trained on computer code, then authoritative and accurate information about software application capabilities can be obtained, but the AI model focuses on the structure of the programming language instead of the capabilities of the software application
Solution Approach 1:
The patent extracts only the relevant functional information from the source code, specifically function calls, their parameters, and return values. By taking out only these critical elements and discarding the surrounding code structure and syntax, the training data focuses exclusively on what the software does rather than how it is implemented in a particular programming language.
Solution Approach 2:
The patent applies different processing quality to different parts of the source code. Rather than treating all code uniformly, it specifically analyzes and extracts functional calls and their semantic meanings, while ignoring syntactic details and language-specific structures. This localized focus ensures the AI learns capabilities rather than programming language patterns.
3Productivity
If existing AI models are used for customer support, then fast and consistent responses can be provided, but human error and lack of authority regarding software-specific capabilities remain
Solution Approach 1:
The patent performs preliminary processing of the source code to create a comprehensive training dataset before the AI model is deployed for customer support. By pre-extracting and formatting all functional information from the source code into training examples, the model is prepared in advance with accurate software capability information, enabling both fast and reliable responses without requiring human intervention during inference.
Data Source
AI summary
Benefits can be obtained from utilizing artificial intelligence (AI) for customer support. However, when the customer support is for a software application, conventional AI-based support bots rely on non-authoritative data sources, rather than the underlying computer code of the software application. While artificial intelligence has been applied to source code, the focus is on the structure of the programming language, rather than what the computer code does or how it works. Accordingly, an AI model is disclosed that can be trained on source code to provide responses about what a software application does, how it works, and/or the like. In an embodiment, source objects implement an AI-supporting interface that can support the AI model in the training and/or operating phases.