docs: update LoCoMo retrieval-only SochDB NVIDIA benchmark

[tatavishnurao]

Summary

This PR updates the LoCoMo retrieval-only benchmark report with a full SochDB-backed NVIDIA retrieval run using metadata-rich memory rendering.

The evaluated configuration uses:

• SochDB vector backend
• NVIDIA llama-nemotron-embed-1b-v2
• 2048-dimensional embeddings
• BM25 + vector hybrid retrieval
• RRF fusion
• single-query retrieval
• metadata-rich memory rendering
• no reranker
• no decomposed retrieval
• no evidence completion

Result

On the full LoCoMo retrieval-only evaluation:

Variant	Hit@100	Recall@100
Raw full K100	82.70	77.46
Metadata full K100	89.53	84.69
Delta	+6.83	+7.24

The strongest gains were in:

Category	Δ Hit@100	Δ Recall@100
multi_hop	+14.61	+15.51
single_hop	+9.96	+12.16
temporal	+7.19	+7.84

Notes

This is a retrieval-only benchmark. It measures gold evidence recovery using Hit@K and Recall@K. It does not report final LoCoMo answer accuracy or judge-based QA accuracy.

Validation

The full metadata run produced:

• 1,986 retrieval rows
• 1,977 scored questions
• 10 LoCoMo samples

Command Configuration

uv run python benchmarks/paper/locomo/runners/run_hybrid_locomo_retrieval.py \
  --memories benchmarks/paper/locomo/data/locomo_memories.jsonl \
  --questions benchmarks/paper/locomo/data/locomo_questions.jsonl \
  --embedding-provider nvidia \
  --embedding-model nvidia/llama-nemotron-embed-1b-v2 \
  --embedding-dim 2048 \
  --vector-backend sochdb \
  --host "$SOCHDB_HOST" \
  --port "$SOCHDB_PORT" \
  --collection-prefix locomo_sochdb_nvidia_dim2048_metadata_full_k100 \
  --k 100 \
  --candidate-k 200 \
  --bm25-weight 1.5 \
  --vector-weight 0.75 \
  --rrf-k 60 \
  --query-mode single \
  --memory-render-mode metadata \
  --retrieval-plan one_shot \
  --evidence-completion none \
  --retrieved-id-mode memory \
  --reranker-provider none \
  --out benchmarks/paper/locomo/results/full_single_metadata_sochdb_nvidia_dim2048_k100/retrieval.jsonl