Introduction
As the ecosystem of Software-as-a-Service (SaaS) platforms continues to expand, the ability to connect these tools into functional, automated workflows—often referred to as "integration flows"—has become increasingly critical. However, end-users face significant barriers when using traditional visual builders like Zapier or Microsoft Power Automate. Specifically, users struggle with discovery (finding the correct trigger or action among thousands of API options) and complexity (mapping data inputs and outputs between disparate systems).
To address these challenges, the Goal Oriented Flow Assistant (GOFA) provides a natural language interface that transforms plain English requests into executable automation workflows. By automating the discovery and composition logic, GOFA allows users to create complex "trigger-action" integrations simply by describing their goals.
The GOFA System Architecture
The GOFA system utilizes a sophisticated three-step pipeline to translate a user's natural language utterance into a functional workflow compatible with the IBM App Connect platform.
Core Methodology
- Abstract Meaning Representation (AMR) Parsing: The system first parses the user's sentence into a rooted, directed graph using AMR. This process abstracts away linguistic noise to identify core tasks, specifically pinpointing the intended connector (the application), the business object (e.g., a "contact"), and the operation (e.g., "create").
- Integration Knowledge Graph (IKG): The IKG serves as a semantic bridge, mapping the tasks identified by the AMR parser to the actual metadata of the target platform. It employs Latent Link Discovery—powered by Graph Neural Networks (GNNs) and embeddings—to resolve naming inconsistencies across different vendors (e.g., recognizing that a "file" in Dropbox is equivalent to a "document" in OneDrive).
- AI Planner (Masterplan): The final stage uses an AI planner to sequence the identified candidates. The planner treats API operations as logical actions with specific prerequisites and effects, ensuring that data dependencies (such as passing an email address from a trigger to a subsequent action) are satisfied. The resulting plan is exported as a YAML file for execution.
Key Insights and Technical Innovations
GOFA introduces several key innovations to ensure the reliability of generated flows:
- Structural Heuristics: The system uses heuristics derived from the AMR graph structure to distinguish between the "trigger" (the event initiating the flow) and the subsequent "actions."
- Latent Semantic Mapping: By using GNNs to capture the relationship of an entity to its surroundings within the knowledge graph, GOFA can resolve cross-vendor terminology discrepancies without requiring manual mapping for every new application.
- Planning as Validation: Unlike keyword-based "guessing," the use of a formal AI planner ensures that the generated workflow is a logically sound sequence where all data requirements are met.
Results and Future Directions
As a functional demonstration, GOFA successfully generates multi-step flows from single-sentence utterances within the App Connect environment. The system provides a visual UI that exposes the AMR parse and IKG mapping, offering transparency into the AI's interpretation process. The final output is a native YAML format that can be directly executed or further refined in a traditional GUI.
Future work will focus on enhancing human-AI collaboration, specifically through robust "repair" features that allow users to correct misinterpretations. Additionally, the authors aim to improve explainability by providing natural language justifications for why specific applications or actions were selected by the planner.
Summary updated