GADR: AI Workflow Generates Architecture Decision Records from Meeting Transcriptions
A recent study presents GADR, a multi-agent workflow that autonomously corrects itself, aimed at extracting architectural decisions from unrefined meeting transcripts and creating drafts of Architecture Decision Records (ADRs) in the Nygard format. This method challenges the prevalent belief in current LLM-based techniques that assumes input data is adequately structured. In reality, architectural decisions frequently arise from informal and chaotic meetings, where choices are often implicit and mixed with unrelated discussions, complicating single-pass prompting. GADR counters this by utilizing multiple agents for iterative output refinement. A feasibility analysis of five actual project meeting transcripts, reviewed by four senior architects and assessed by fifteen students, indicated that GADR effectively captures most expert-identified decisions and generates drafts deemed clear and useful by participants. Additionally, it surpassed zero-shot and few-shot benchmarks in terms of stability and structural integrity, indicating a stronger approach for generating ADRs from unstructured discussions.
Key facts
- GADR is a multi-agent, self-correcting workflow for extracting architectural decisions.
- It generates Nygard-formatted Architecture Decision Record (ADR) drafts.
- Existing LLM-based approaches assume input is already structured.
- Architectural decisions often emerge from informal, noisy meetings.
- Feasibility study used five real project meeting transcripts.
- Four senior architects provided expert review.
- Fifteen students evaluated the generated drafts.
- GADR outperformed zero-shot and few-shot baselines in stability and structural quality.
Entities
—