Category-Scoped Retrieval for RFP Response Generation: Design and Evaluation of a Category Schema Aligning Requirement Extraction and Corpus Indexing in a RAG System

Sammanfattning

Responding to Requests for Proposals (RFPs) requires consultants to manually work on two sides. The first is extracting requirements from the RFP. The second is finding and reusing past proposal content to help answer them. This challenge is also apparent at the Swedish consulting firm operating in the data, cloud, and AI sector for which this thesis was conducted. Retrieval-Augmented Generation (RAG) could automate this work, but reviewed designs treat both sides as separate problems. Since both sides share a similar topical structure, aligning their categories could focus retrieval on the most relevant content. This thesis tests that idea. Using Design Science Research, it builds a RAG prototype that connects both sides through a shared category schema derived from the firm's own RFPs and past proposals. Two retrieval configurations are compared on three real RFPs and rated by six consultants. The first filters by category before searching. The second searches the full corpus. The thesis delivers two outputs. The first is the prototype. It addresses the firm's practical challenge of automating content reuse. The second follows from evaluating the prototype on the company's corporate data. It is design knowledge about when category scoping helps. The overall results show it did not. Retrieval precision differed by less than one percentage point between both configurations. However, when combining this result with the finding that 88.9% of retrieved content was identical, a condition can be derived. It lies in the system architecture, where retrieval consists of two components, a category filter and similarity search. The similarity search compares two LLM-generated descriptions sharing a similar, rich vocabulary. Because these categorical differences are already covered by the descriptions themselves, category filtering becomes redundant. Three further candidate conditions are derived from category-level results and reported as hypotheses for future work, as the data is not sufficient to confirm them.