Better retrieval didn't make the answers more faithful. It made the model willing to answer.

Korean is a morphologically rich, agglutinative language. Particles attach to nouns, verbs inflect heavily, and whitespace does not mark word boundaries the way it does in English. That makes it a good place to test whether retrieval methods built on English assumptions hold up.

I used Mr. TYDI Korean, a standard multilingual retrieval benchmark, and compared four retrievers on the same queries. BM25 with naive whitespace tokenization. BM25 with Kiwi, a Korean morphological analyzer, keeping only content morphemes. Dense retrieval with multilingual-e5-base. And an equal-weight reciprocal-rank fusion of the Kiwi and dense runs.

Retrieval metrics on their own only tell you what reached the context window. The second half of the study ran all four retrievers through the same generator on a 200-query subset, holding the model constant so that any difference in answer quality traced to retrieval rather than to the generator. A separate judge scored each answer for faithfulness and relevance, and I tracked how often the model declined to answer at all.

The tokenizer was worth more than I thought

Swapping whitespace tokenization for Kiwi lifted Recall@10 from 0.213 to 0.311 and Recall@100 from 0.306 to 0.502. Nothing else changed. That is a large gain for a one-line substitution, and it means a lot of published Korean BM25 baselines are probably measuring their tokenizer rather than their retriever.

I got the opposite result to the benchmark's own paper

Mr. TYDI reported that BM25 outperformed multilingual DPR on Korean. With multilingual-E5 I found the reverse, and by a wide margin: 0.658 Recall@10 against 0.311 for a properly tokenized BM25, roughly 2.1 times better. The original finding appears to depend heavily on which dense model was available at the time rather than on anything intrinsic to Korean. That is worth knowing before you reach for a sparse baseline on the strength of a 2021 result.

The hybrid made things worse

I expected reciprocal-rank fusion to win. It usually does. Here it landed at 0.478 Recall@10, well below dense alone at 0.658, while holding up at depth (0.810 Recall@100 against 0.829). Equal-weight fusion pulled strong dense results down the ranking to make room for weaker BM25 ones. When one retriever substantially dominates the other, naive fusion costs you precision at the top, which is the part that reaches the model.

And the one that mattered most

I expected better retrieval to produce more grounded answers. It did not. Faithfulness scores sat between 4.625 and 4.780 on a five-point scale across all four retrievers, effectively flat. What moved was whether the model answered at all. With whitespace BM25 it declined 57.5% of the time. With dense E5 that fell to 18.5%.

That behavior is the thing worth probing further, because it was measured entirely inside a grounded pipeline. Every condition had retrieval, some of it just bad. The obvious next question is what happens when there is no retrieval at all.

The follow-up moves to Korean corporate filings from DART, the Korean regulator's disclosure system. Four conditions, one variable:

Two design choices carry the weight. The first is that the grounded system and the ablation share a generator, so the delta between them isolates retrieval rather than model quality. The second is the corpus window. Filings from 2024 and 2025 sit at or past the models' training cutoffs, so a raw model has to either abstain or invent, which is the behavior under test. Wikipedia would have been useless here: frontier models already know most of it, and the grounding signal collapses.

Raw models are explicitly allowed to say 확실하지 않습니다 (“I’m not certain”). Forcing an answer turns every uncertainty into a fabrication and makes the hallucination metric meaningless.

I could be wrong about that, which is the point of writing it down first. The alternative outcome is that the frontier models abstain cleanly on recent filings they have no memory of, in which case the value of grounding is narrower than I think and the interesting number becomes how much accuracy it recovers rather than how much fabrication it prevents.