Retrieval reads a sample. Some questions need every row.
Ten Larkspur Cloud server sizes, one Document per row, ranked by real all-MiniLM-L6-v2 cosine similarity. Move k and compare the best answer a reader could give from the rows above the magenta rule with the answer a routed query gives from all ten.
- Rows the model reads
- Best answer from those rows
- Routed answer, all ten rows
- Retrieval answer right
Rows ranked for this question
Cosine similarity, all-MiniLM-L6-v2
| Rank | Server and the text embedded | Score | Price | Needed |
|---|
Scores computed offline with sentence-transformers and checked against a LangChain FAISS index (same order for all eight question and text pairs). The context the model reads carries each row’s name and price from metadata, as in the lesson’s format_docs. “Best answer” is what a careful reader could compute from those rows; it is arithmetic, not a model output.
All four questions at this k
| Question | Route | From top-k | Routed |
|---|
The routed column assumes the router labels each question correctly, as the lesson’s router does for these four kinds.