Abstract
This disclosure describes techniques to enhance the efficacy of root cause analyses in software defect resolution by integrating similarity models into a ReAct (reasoning and acting) framework. The ReAct framework, which combines reasoning using a large language model (LLM) with tool-assisted actions, can suffer from hallucinations or infinite loops. To address this, a similarity advisor model is introduced that uses unsupervised clustering to group previously resolved issues based on textual problem descriptions and associated artifacts such as error logs. When the ReAct agent enters unproductive cycles during the resolution of an issue, the similarity advisor model clusters the issue into an existing cluster, potentially revealing cluster metadata that includes steps to resolve the issue. The LLM uses resolution steps from past successes and contextualizes them to the issue on hand. The resulting feedback loop can significantly improve the probability of successful triaging of software defects by learning from both human and AI-led resolutions.
Creative Commons License
This work is licensed under a Creative Commons Attribution 4.0 License.
Recommended Citation
Elumalai, Babu Prasad; Aggarwal, Garvit; Rana, Shreya; Jain, Aneesha; and Khurana, Shubham, "Applying Similarity Detection to Improve Success Probability for Root Causing Issues Within a ReAct Framework", Technical Disclosure Commons, (May 30, 2025)
https://www.tdcommons.org/dpubs_series/8174