https://www.mdu.se/

Architectural decisions shape system quality, yet the design space is too large to explore exhaustively, so architects often fall back on familiar patterns and weigh few alternatives. Large language models can propose designs but lack organizational context, and the knowledge that would supply it lies scattered across decision records, documentation, and code.

This thesis investigates how retrieval-augmented large language models can support the generation and iterative refinement of context-aware architectural alternatives.

We propose a workflow that structures a problem description into a decision frame, composes a retrieval query from it, and retrieves from wiki pages, architecture decision records (ADRs), and a code graph. Two alternatives are generated, each with its reasoning, trade-offs, risks, and sources. The architect refines them by excluding one with a stated reason or by providing guidance, and the accepted alternative is published as an ADR and ingested back into the knowledge base. The prototype was evaluated over five scenarios using automated quantitative metrics and qualitatively with a software architect.

Retrieval reached a mean context precision of 0.921, and context recall was perfect in four of the five scenarios. Faithfulness differed between the two alternatives in every scenario, indicating that generation exposed distinct design directions rather than converging on one.

In conclusion, our work presents an approach for generating and iteratively refining architectural alternatives that are grounded in organizational knowledge and captured as reusable decision records.