The old BM25 does not retreat even in front of Khmer. In the KSE-Web study, a research group compared character n-gram lexical retrieval, multilingual dense retrieval, a hybrid combination, and query expansion assisted by Qwen2.5 models. On a dataset built from roughly 17,000 candidate titles and filtered down to 3,000 full-text Khmer documents, with 300 manually reviewed user-style queries and partially human-verified silver relevance labels, the result is clear: BM25 reaches 0.943 Recall and 0.876 nDCG. The hybrid approach comes close (0.929 Recall and 0.871 nDCG), while dense retrieval alone lags behind.

The most interesting finding concerns LLM-assisted expansion. It not only fails to outperform non-expanded retrieval, but it also introduces topic drift, generic terms, and noisy reformulations. There is, however, a meaningful difference between the two models used: Qwen2.5-3B produces substantially better expansions than Qwen2.5-0.5B. This suggests that for a low-resource language like Khmer, model size and expansion quality matter, but they are not enough to overturn the final ranking. The study also observes that simple filtering of expansions can remove useful semantic cues. In low-resource languages, over-cleaning generated text risks discarding precious synonyms or spelling variants, while under-cleaning lets noise through.

Behind the technical comparison lies a structural lesson for teams designing local search stacks. Khmer has ambiguous word boundaries, limited annotated data, weak support in multilingual embedding models, and frequent mixing with English. In this context, the robustness of character n-grams acts as an anchor: it requires no GPU, does not depend on a cloud service, and keeps operational costs low. The hybrid with dense retrieval adds little in nDCG and introduces complexity. For on-premise deployments, the point is not that the LLM is useless, but that it must be placed deliberately: if expansion does not improve retrieval, adding even a small model risks consuming resources without measurable benefit.

There is also a data quality problem. Silver labels with only partial human verification make the benchmark less solid than the numbers might suggest. Anyone using this work as a foundation will need to invest in more extensive human annotations, because Khmer-aware models do not emerge on their own: they come from verified datasets. The study explicitly recognizes this and points precisely there for future work.

For vendors and teams working on low-resource languages, the result is a reminder: multilingual embedding models are not a universal shortcut. And for those evaluating self-hosted retrieval pipelines, the question is not only which LLM to run locally, but whether the LLM has a real role in the retrieval stage. In this case, BM25 character n-grams continue to do the heavy lifting, while generative AI shows clear limits when the language is not well represented in pre-training. The promised GitHub repository is a first step; the robustness of the conclusions will depend on how much the community can strengthen human verification.