locomo: add grep-allviews K200 retrieval results (best K=20/50 recall)

[tatavishnurao]

Publishable Grep+Allviews Retrieval Pipeline (K=200)

Single-pass retrieval with 4 multiviews, BM25+SochDB vector+grep (trigram) RRF fusion, and local neighbor expansion.

Results

K=20

Category	n	Hit@20	Recall@20
adversarial	446	0.904	0.901
multi_hop	89	0.730	0.628
open_domain	841	0.954	0.946
single_hop	281	0.918	0.670
temporal	320	0.934	0.912
overall	1977	0.924	0.877

K=50

Category	n	Hit@50	Recall@50
adversarial	446	0.957	0.955
multi_hop	89	0.854	0.737
open_domain	841	0.985	0.982
single_hop	281	0.968	0.819
temporal	320	0.975	0.965
overall	1977	0.969	0.939

K=100

Category	n	Hit@100	Recall@100
adversarial	446	0.980	0.978
multi_hop	89	0.899	0.795
open_domain	841	0.994	0.993
single_hop	281	0.996	0.903
temporal	320	0.994	0.984
overall	1977	0.987	0.966

K=200

Category	n	Hit@200	Recall@200
adversarial	446	0.993	0.992
multi_hop	89	0.921	0.858
open_domain	841	0.995	0.995
single_hop	281	1.000	0.961
temporal	320	0.997	0.996
overall	1977	0.992	0.984

Pipeline Config

• 4 multiviews: turn + event + entity + neighbor_window
• BM25 + SochDB vector + Grep (trigram) RRF fusion (bm25=0.1, vec=3.0, grep=0.5, rrf_k=60)
• Local neighbor expansion: ±2 turns from top-50 anchors
• NVIDIA llama-nemotron-embed-1b-v2 (2048-dim)
• K=200, candidate_k=400, single query mode