Latest LoCoMo-retrieval only scores across @10 @20 @200 achievement & its artifact

[tatavishnurao]

Summary

This PR adds benchmark artifacts for the full_topk_rerank LoCoMo retrieval run.

The run demonstrates a strong retrieval-only ceiling using SochDB candidate retrieval followed by full top-k cross-encoder reranking over the LoCoMo benchmark set.

This result differs from all the previous runs as the individual categorical scores every category on k=200 hits 90 above including multi-hops score to be @hit = 0.966292 and @recall = 0.916063. This is truly an achievements with the latest PR changes done on the ongoing build that focus on benchmark retrieval sake. This does not misalign with the principles of SochDB.

Benchmark Pipeline

The numbers come from a retrieval-only pipeline:

Official LoCoMo memories + questions
        ↓
SochDB retrieval candidate generation
        ↓
metadata rendering + BM25 / vector fusion
        ↓
top-50 candidate window
        ↓
BAAI/bge-reranker-v2-m3 cross-encoder reranking on CUDA
        ↓
full reranked retrieval artifact
        ↓
official retrieval scorer
        ↓
Hit@K / Recall@K report

Key Results

K	Overall Hit	Overall Recall
10	0.847749	0.795667
20	0.915023	0.868808
50	0.950430	0.917037
100	0.975721	0.952834
150	0.987355	0.972464
200	0.990389	0.981253

Category Results

K=20

Category	n	Hit	Recall
adversarial	446	0.919283	0.918161
multi_hop	89	0.696629	0.592146
open_domain	841	0.929845	0.921720
single_hop	281	0.925267	0.692227
temporal	320	0.921875	0.892969

K=100

Category	n	Hit	Recall
adversarial	446	0.957399	0.956278
multi_hop	89	0.921348	0.823460
open_domain	841	0.980975	0.977011
single_hop	281	1.000000	0.897564
temporal	320	0.981250	0.969010

K=200

Category	n	Hit	Recall
adversarial	446	0.979821	0.978700
multi_hop	89	0.966292	0.916063
open_domain	841	0.994055	0.993064
single_hop	281	1.000000	0.959638
temporal	320	0.993750	0.990885

Note to onboarding developers that this is purely retrieval only scores and truly reproducible by anyone on the build verison: 2.0.12