RAG research digest: config failures, poisoned HNSW indexes and agent memory

A roundup of arXiv papers from October 12–19, 2026 on production RAG issues: a 654-case study of configuration failures, a topology-aware attack hitting 31% success on a 2M-document HNSW index, a 2.2–2.7x GraphRAG speedup, and the FEB-30 benchmark for agent memory.

DEV Community

Open source

RAG research digest: config failures, poisoned HNSW indexes and agent memory

A 654-case study across 10 RAG systems yielded 9 actionable configuration fixes
Topology-aware poisoning of HNSW indexes reached 31% success on 2M Wikipedia documents
FuseRAG cut GraphRAG time-to-first-token by 2.2–2.7x with no quality loss
FEB-30 shows memory systems classify memory operations correctly only 20% of the time

#GraphRAG#HNSW#FAISS#Milvus

MLPerf Inference v6.1 adds end-to-end RAG and desktop agent benchmarks

laya-compactor cuts RAG context tokens by 70% with no answer-quality loss

Nozzle: Google is poisoning search results for AI scrapers and rank trackers

Software

RAG research digest: config failures, poisoned HNSW indexes and agent memory

MLPerf Inference v6.1 adds end-to-end RAG and desktop agent benchmarks

laya-compactor cuts RAG context tokens by 70% with no answer-quality loss

Nozzle: Google is poisoning search results for AI scrapers and rank trackers

Pinterest's Manas Platform Evolves from HNSW to Quantized SPANN