Highnet Safe LLM Tools

Sheet 2 ·How well does it work?

The passage holding the answer was in the top 5 for 84% to 99% of 105 questions, depending on the settings; weakest: BM25 · small.

Does the search find the answer?

recall@5: the share of 105 answerable questions whose answer-holding passage was among the top 5 retrieved. The model answers only from what it is given, so this caps everything after.

MRR rewards finding it early: 1 for the top passage, ½ for second, 0 outside the top 10.

Measure
Small chunks
SearchReranker offReranker on
BM25recall@5 84%, over 105 questionsrecall@5 97%, over 105 questions
Vectorrecall@5 89%, over 105 questionsrecall@5 98%, over 105 questions
Hybridrecall@5 92%, over 105 questionsrecall@5 99%, over 105 questions
Medium chunks
SearchReranker offReranker on
BM25recall@5 93%, over 105 questionsrecall@5 99%, over 105 questions
Vectorrecall@5 90%, over 105 questionsrecall@5 99%, over 105 questions
Hybridrecall@5 96%, over 105 questionsrecall@5 99%, over 105 questions
Large chunks
SearchReranker offReranker on
BM25recall@5 94%, over 105 questionsrecall@5 99%, over 105 questions
Vectorrecall@5 91%, over 105 questionsrecall@5 98%, over 105 questions
Hybridrecall@5 97%, over 105 questionsrecall@5 99%, over 105 questions