Graph-of-thought Framework
Overview
The Graph of Thought (GoT) framework is under active research at Sully AI as a foundation for structured, multi-step reasoning across complex clinical workflows. GoT replaces traditional linear logic with a dynamic graph architecture, where nodes represent individual facts, hypotheses, or actions, and edges encode relationships such as causality, temporal ordering, or contradiction.
This system is not yet deployed, but initial planning has focused on integrating GoT into our decision support stack and future consensus mechanism. Team discussions have explored mapping GoT nodes to FHIR-compatible data structures and using it to coordinate multi-agent reasoning in diagnosis, triage, and task planning.
Why Graph of Thought?
Supports Multi-Agent Coordination
Agents can independently operate on subgraphs, contribute to shared nodes, or prune irrelevant branches.
Ideal for collaborative reasoning between scribe, diagnosis, and referral agents.
Improved Reasoning Structure
Enables branching logic, revisitation of earlier steps, and parallel evaluation paths.
Useful for differential diagnosis, treatment option evaluation, and ambiguous data interpretation.
Better Alignment with Clinical Workflows
Reflects how clinicians think: weighing evidence, comparing timelines, revisiting prior assumptions.
Facilitates integration with structured clinical data (e.g., FHIR, SNOMED, LOINC).
Composability and Reusability
Reasoning fragments can be reused across visits, patients, or similar cases.
Ideal for building longitudinal patient reasoning models.
Explainability and Auditability
Every conclusion can be traced to its contributing nodes and agent decisions.
Graph structure enables clear visualization of reasoning paths and confidence levels.
Planned Integration Points
Consensus Engine: Graph merge strategies across agents for unified decision outputs.
Temporal Patient Modeling: Using GoT to build case timelines (e.g., symptom onset → lab results → resolution).
EHR Interoperability: Mapping graph nodes to FHIR resources and external knowledge graphs.
Uncertainty Management: Representing ambiguous or competing conclusions as competing branches in the graph.
Last updated