OKTS retrieval evaluation -- raw output
============================================================
tokenizer:  tiktoken cl100k_base
corpus:     148 tools, 20 servers, 91 categories
queries:    41 labeled cases   k=5

retriever                       acc@1  acc@k    MRR   coll     tok
------------------------------------------------------------------
Flat BM25 (baseline)           92.7% 100.0%  0.963  95.1%     532
Graph-aware (full)             97.6% 100.0%  0.988 100.0%     531

ablation:
------------------------------------------------------------------
BM25 only (baseline)           92.7% 100.0%  0.963  95.1%     532
Dense only (hashing)           92.7% 100.0%  0.957 100.0%     530
Hybrid, no structure           97.6% 100.0%  0.988 100.0%     531
BM25 + hierarchy               92.7% 100.0%  0.963  95.1%     532
BM25 + graph                   92.7% 100.0%  0.963  95.1%     530
Hybrid + hierarchy + graph     97.6% 100.0%  0.988 100.0%     531

token reduction sweep:
     N       raw      okts   reduction
--------------------------------------
    11       618       526       14.9%
    25      1408       526       62.6%
    50      2870       521       81.9%
    75      4634       528       88.6%
   100      6193       529       91.5%
   125      7683       531       93.1%
   148      9041       531       94.1%
