Comparison
Recall@kvsTop-k retrieval
Recall@k
build a 50-question set with known source chunks and measure recall before touching generation; most bad RAG is bad retrieval.
The fraction of test queries whose correct chunk appears in the top k results: the core retrieval metric, measurable without any LLM. If recall@k is low, no prompt engineering downstream can save the answers.
Full entry →Top-k retrieval
you raised k to twenty, the right chunk was definitely in there, and the model missed it among the noise anyway.
Returning the k best chunks, k typically 3 to 20. Small k risks missing the answer; large k pads the context with noise, cost and lost-in-the-middle risk. Tune it with an eval set, not by feel.
Full entry →