2. Thematic Analysis of Retrieval-Augmented Approaches
Section excerptThe first wave of retrieval-augmented generation systems treated retrieval as a one-shot upstream step, concatenating top-k passages into a fixed context window before generation[1]. Subsequent work relaxes this rigidity: iterative retrievers re-query the corpus after each generation step[3], while self-reflective variants emit explicit retrieval tokens to decide when additional context is warranted[7].
For long-form scientific writing the trade-off is sharper. Static retrieval is fast and reproducible but tends to over-represent the first few sub-topics; iterative schemes adapt but compound latency and citation drift[4, 9]. A growing body of evidence suggests that the highest-quality surveys come not from richer prompting but from better candidate sets: composite scoring functions over relevance, citation count, and recency consistently outperform pure semantic ranking on coverage and novelty[12, 14].
… continues for five more sections — methodology comparison, key findings, gaps, and conclusion.