RAG retrieval quality, measured

Three retrieval configurations answer the same questions over the same Kubernetes documentation snapshot, with the same answer model and prompt. Pick a question to compare what each one retrieved, what it answered and how it scored.

Measured From benchmark run 20261105-0038-d71e. Question type: exact_term. Review status: accepted.

Baseline

Fixed-size chunks, dense search

NOT_IN_DOCS: The passages do not state what that exact command retrieves.

Answer correctness
incorrect
Hit@5
no
Recall@5
0.00
MRR@10
0.11

2849 ms total · cost not priced

Intermediate

Structure-aware chunks, dense search

NOT_IN_DOCS: The passages do not explain what this exact `curl` command retrieves.

Answer correctness
incorrect
Hit@5
no
Recall@5
0.00
MRR@10
0.17

3486 ms total · cost not priced

Advanced

Structure-aware chunks, hybrid search, RRF, rerank

It retrieves a list of Pods in the `default` namespace, returned as a `PodList`.

Answer correctness
correct
Hit@5
yes
Recall@5
1.00
MRR@10
1.00

131397 ms total · cost not priced