Several phrasings, one fused ranking.
The twelve Larkspur Cloud chunks and TF-IDF from the M10 lab. Each query variant retrieves its own top 4; RRF or a flat union merges the lists. The chunks above the magenta rule are what the model reads.
- Recall@n, variants
- Recall@n, original question
- Mean recall@n, all three
- Retriever calls
Merged ranking
| # | Chunk | Score | Rank in each variant | Golden set | Text |
|---|
Ranks count from 0, as enumerate does in the 03a notebook; a hyphen means that variant did not retrieve the chunk. Ties keep the order in which the chunks were first seen.
Recall at this setting
| Golden question | Original | Variants |
|---|---|---|
| Mean |
The Original column is the M10 baseline: the question as asked, ranked by the same TF-IDF index. At two chunks it reads 0.50, the M10 lab’s baseline row. The variants are representative, not model output; the retrieval, fusion and recall are computed here.