Source-linked AI summary
GADR: Gathering Architecture Decision Records from Meeting Transcriptions
Lucas Daniel Costa da Silva, Kiev Gama
TL;DR
GADR addresses the limited support for generating ADRs from noisy, unstructured meeting transcripts. Its multi-agent workflow produced stable, structured, and generally useful drafts, while requiring human review because retrieval sometimes introduced unsupported details.
Problem
Existing ADR-generation approaches largely assume reasonably structured input, despite architectural decisions often emerging from long, noisy, fragmented conversations.
Method
GADR uses a multi-agent, self-correcting finite-state workflow with retrieval augmentation to extract decisions from meeting transcripts and generate Nygard-formatted ADR drafts.
Results
The agentic workflow produced more stable, structurally adherent ADRs, captured most expert-identified decisions, and generated drafts that participants generally found clear and useful.
Takeaways & Limitations
GADR is best positioned as a semi-automated documentation assistant that produces reviewable ADR drafts rather than authoritative architectural records.
Takeaways & Limitations
Retrieved context sometimes introduced unsupported details, so generated ADRs require stakeholder review to verify transcript faithfulness and decision status.
Abstract
from arXiv · showhide
Existing LLM-based approaches to Architecture Decision Record (ADR) generation share a critical and largely unexamined assumption: that input is already reasonably structured. In practice, architectural decisions emerge from informal, noisy meetings where choices are implicit, fragmented, and entangled with off-topic dialogue, precisely the conditions under which single-pass prompting degrades. This paper presents GADR, a multi-agent, self-correcting workflow that extracts architectural decisions from raw meeting transcriptions and generates Nygard-formatted ADR drafts. A feasibility study comprising five real project meeting transcripts, expert review by four senior architects, and evaluation by fifteen students provides initial evidence that the agentic workflow captures most expert-identified decisions and produces drafts participants found clear and useful, outperforming zero-shot and few-shot baselines in stability and structural adherence. The study also addresses the underexplored trade-off of RAG-based enrichment improving ADR depth while simultaneously risking transcript-unfaithful content, raising open questions about traceability in automated architectural documentation that we believe is worth the community's attention.
1 Introduction
Architecture Decision Records preserve the historical context of decisions, but their adoption is hindered by documentation effort and uncertainty about what to capture. GADR addresses noisy meeting transcripts with a multi-agent workflow that extracts, critiques, and refines decisions into Nygard-formatted ADR drafts.
- ADRs document architectural decisions and support Architectural Knowledge Management by preserving the historical context of decisions made by stakeholders over time.
- Adopting ADRs is difficult because capturing Architectural Knowledge requires significant effort and teams lack clarity about which elements warrant documentation.
- Fragmented meeting dialogue mixed with off-topic content causes zero-shot and few-shot prompting to suffer attention degradation, disorganized outputs, and reduced accuracy.
- GADR introduces a multi-agent, self-correcting workflow that extracts, critiques, and refines decisions from raw meeting transcriptions into Nygard-formatted ADR drafts.
2 Background And Related Work
Architectural decision-making is often fragmented for students and novice architects, while LLM-based ADR generation remains challenged by long, noisy conversational inputs. These limitations motivate multi-agent workflows that decompose, coordinate, and validate ADR-generation tasks.
- Motivation: Students and novice architects often make fragmented, non-linear decisions while transitioning from problem to solution spaces.They may struggle to identify suitable concepts, compare technological alternatives, and build consensus.
- Motivation: Inexperienced engineers may favor familiar technologies, processes, or approaches instead of exploring alternatives, potentially leading to misguided choices.Human and operational factors strongly influence their decision-making.
- ADR Generation: ADRs document decisions, context, rationale, and consequences, while zero-shot, few-shot, fine-tuning, and RAG support LLM-based ADR generation from textual context.These approaches emerged partly because ADR adoption faces practical barriers.
- Problem: Long and noisy conversational inputs can distract LLMs and reduce generation accuracy, motivating multi-agent extraction of architectural decisions from meeting transcriptions.Excessive or poorly selected context is identified as a specific challenge.
- Multi-Agent Orchestration: Graph-based and multi-agent architectures decompose complex tasks into specialized roles, while structured state control and review patterns support adaptable workflows and factual accuracy.Examples include MetaGPT, Planner-Executor, and Critic-Reviewer designs.
3 Proposed Approach
GADR extracts architectural decisions from noisy meeting transcripts and generates structured ADR drafts through a finite-state, multi-agent architecture augmented with RAG. Its agents coordinate via shared mutable state, bounded workflow nodes, and iterative critique to refine extracted decisions.
- Architecture: GADR replaces linear single-pass processing with a finite-state, multi-agent architecture augmented by Retrieval-Augmented Generation.The system automatically extracts architectural decisions from meeting transcriptions and generates structured ADR drafts.
- State coordination: A centralized mutable state S lets LangGraph agents update context, audit reasoning steps, and support conditional routing and looping.Nodes read and update shared contextual variables rather than relying on sequential prompt chaining.
- Workflow components: Workflow nodes perform bounded operations including LLM calls, critique, retrieval, and formatting while coordinating local modules with external services.Project artifacts and ChromaDB retrieval remain local, while Gemini API reasoning and Tavily web search are external services.
- Decision extraction: The extraction node identifies decisions, rationale, and rejected alternatives, then uses critique vectors to merge overlaps, split tangled concerns, and remove noise.The first iteration frames the model as a senior software architect; later iterations refine the previous output.
- Retrieval enrichment: RAG enriches architectural explanations, but retrieved content remains subject to human review.The approach explicitly treats retrieved information as a potential limitation requiring verification.
4 Methodology And Setup
The study evaluates a novel multi-agent artifact for extracting and documenting architectural decisions from noisy meeting transcripts through an exploratory mixed-method Engineering Research design. It combines comparisons with zero-shot and few-shot prompting, expert review, student questionnaires, and a reference ADR dataset used for RAG and semantic validation.
- Study goal: The primary goal is to analyze the multi-agent system’s effectiveness in extracting and documenting ADRs from raw meeting transcripts.
- Research design: The Engineering Research study designed, implemented, and empirically evaluated a software engineering artifact in an educational software architecture setting.The evaluation used an exploratory mixed-method design.
- Data and reference materials: The study uses ADRs from the Buchgeher et al. MSR dataset as both the RAG module’s primary knowledge base and a gold-standard source of semantic examples.
- Meeting transcripts: The meeting data comprise five architecture kick-off transcripts from undergraduate student teams and one senior-researcher team, characterized by informal language, transcription errors, overlapping dialogue, and fragmented rationale.
- Comparative evaluation: The evaluation compares zero-shot prompting, few-shot prompting with static ADR examples, and a multi-agent workflow implemented as a multi-node directed graph.All three setups processed the transcripts uniformly; the zero-shot condition used a direct Nygard-standard prompt, while few-shot used examples from the reference dataset.
- Human evaluation: Human-centered evaluation combines five-point Likert questionnaires from consenting transcript-providing teams with reviews by four senior software architects who identified core decisions as an approximate ground truth.Questionnaires assessed agreement with extracted decisions, clarity, completeness, and perceived educational value, with open-ended justification.
5 Results
Across preliminary evaluations, GADR produced more stable, structurally consistent, and richer ADRs than static prompting baselines, while retrieval introduced risks of transcript-unfaithful content. Expert and participant assessments indicated broad decision coverage and positive perceptions, alongside omissions and overinterpretation failures.
- Prompting comparison: GADR produced more stable ADR sets than zero-shot and few-shot prompting, whose outputs varied in decision grouping and sometimes merged unrelated decisions or passing comments.Its critique stage reviewed candidate decisions before document generation, reducing output variation.
- Prompting comparison: GADR enriched sparse transcripts with architectural terminology and external context, producing richer explanations of trade-offs, alternatives, and long-term consequences.The added verbosity was considered useful in the study’s educational context because novice students often discuss decisions informally.
- Failure modes: Retrieval introduced contextual contamination, including unsupported claims such as “500 concurrent connections and 5000+ Daily Active Users (DAU)” and “93% industry adoption”.Traceability inspection attributed the first fragment to a retrieved external ADR rather than the meeting transcript; other failures included turning tentative discussions into accepted decisions and omitting lower-granularity decisions.
- Participant evaluation: 52 of 55 evaluations agreed or strongly agreed with the extracted decision, while 51 of 55 found the ADRs helpful for remembering decisions and clear and complete.The participant evaluation included 15 respondents from five student groups and produced 55 decision-level evaluations.
- Expert evaluation: Approximately 18 of the 23 senior decisions were covered by at least one generated ADR, indicating broad alignment with expert-identified architectural concerns.The agent generated 17 ADRs across five transcripts, while seniors listed 23 reference decisions.
6 Discussion
The discussion presents GADR as a semi-automated decision-support workflow that improves stability, ADR structure, and draft usefulness when processing noisy meeting transcripts. Its outputs remain reviewable drafts because extraction errors and retrieval-based enrichment can compromise transcript faithfulness.
- GADR is better understood as decision support than as a fully autonomous replacement for human architectural documentation.Decomposing extraction, critique, refinement, retrieval, and formatting reduces instability from processing noisy transcripts in a single prompt.
- 18 of 23 expert-identified decisions were covered, indicating effective extraction but not full accuracy because omissions and tentative-versus-accepted confusion remained.51 of 55 evaluations considered the ADRs clear and complete, supporting their use as reviewable drafts rather than final records.
- Student participants recognized most extracted decisions and reported that the ADRs improved memory of discussions and understanding of architectural documentation.The educational benefit is relevant because novice architects often discuss decisions informally without separating context, alternatives, consequences, and rationale.
- Agentic RAG produced more detailed ADRs, but retrieved external context introduced unsupported metrics, statistical claims, and stronger-than-intended architectural framing.The discussion therefore treats enrichment as useful but potentially contaminating when it adds content not grounded in the transcript.
- Expert review found that GADR captured most central concerns but sometimes promoted tentative ideas, omitted fine-grained concerns, or merged related choices into broader ADRs.Future versions should expose decision status and granularity so users can inspect, split, merge, or reclassify candidates.
- Generated artifacts should remain reviewable drafts, with stakeholders correcting unsupported extrapolations, confirming decision status, and selecting which ADRs enter the official record.GADR reduces the effort of transforming informal meetings into readable drafts and provides educational value for students learning architectural reasoning.
7 Threats To Validity
The agentic workflow improved ADR structure and stability over single-pass prompting but introduced validity risks, including decision overstatement, omitted concerns, unsupported retrieved details, and potentially obscuring verbosity.
- The workflow sometimes transformed tentative discussions into accepted decisions, risking overstatement of architectural certainty.
- It could omit fine-grained concerns or introduce unsupported details from retrieved external context.
- Verbose ADRs may help novice architects learn, but can obscure which information was actually discussed.
8 Conclusion
GADR presents an agentic approach for extracting architectural decisions from meeting transcriptions and generating ADR drafts, with preliminary evaluation against zero-shot and few-shot prompting. The study argues that LLM-based ADR generation should remain human-in-the-loop while shifting documentation effort toward review.
- GADR extracts architectural decisions from meeting transcriptions and generates ADR drafts through an agentic approach.The approach addresses the effort required to capture architectural knowledge in practice.
- The preliminary evaluation compares GADR with zero-shot and few-shot prompting.
- LLM-based ADR generation should remain human-in-the-loop, with documentation effort shifting toward reviewing generated drafts.The passage frames review as lighter than writing documentation.