Listen to this story
|
JPMorgan has introduced FlowMind, a novel approach leveraging LLMs to create an automatic workflow generation system.
The field of Robotic Process Automation (RPA) has made significant advancements in automating repetitive tasks. However, its effectiveness diminishes when faced with spontaneous or unpredictable user demands.
FlowMind utilises a generic prompt recipe for a lecture, grounding LLM reasoning with reliable APIs. This approach not only mitigates hallucinations in LLMs but also ensures data integrity and confidentiality by eliminating direct interaction between LLMs and proprietary data or code.
FlowMind simplifies user interaction by presenting high-level descriptions of auto-generated workflows, enabling effective inspection and feedback. FlowMind significantly outperformed the GPT-Context-Retrieval baseline method, even without user feedback.
Additionally, the paper introduces NCEN-QA, a new dataset in finance for benchmarking question-answering tasks from N-CEN reports on funds. Researchers evaluated the performance of workflows generated by FlowMind using NCEN-QA and demonstrate its success, the importance of each lecture component, and the effectiveness of user interaction and feedback.
The FlowMind framework operates in two primary stages:
- Lecture to LLM: Providing the LLM with context, available APIs, and the need for workflow generation.
- Workflow Generation and Execution: Utilising APIs to generate workflows and deliver results to users, with an optional feedback loop for user interaction.
Future work may explore crowdsourcing user feedback to refine workflows at scale and life-long learning over past user-approved examples to evolve FlowMind’s performance over time. Additionally, FlowMind could be expanded to handle large libraries of APIs by retrieving the most relevant APIs for a given task based on embedding similarity.