Killing RAG: How a Six-Line Rule Replaces Retrieval-Augmented Generation as Memory
BAAI
An AI researcher argues that RAG (Retrieval-Augmented Generation) is fundamentally flawed as a memory mechanism because vector similarity does not equal temporal truth. They propose a six-line rule that prioritizes the most recent unfogotten version of a fact, treating time as a first-class entity. The rule, implemented in a Rust system, demonstrates higher accuracy than a two-stage neural pipeline with bge-m3 and gemma3 on counterfactual tests.
The author claims to have 'killed' RAG as a memory approach, not as a search method, by identifying a substitution: RAG answers 'what is similar' but not 'what is true now'. Using an example of a person with three job titles (engineer, group leader, director), they show that vector similarity treats all versions equally, leading to hallucinations. The solution is a six-line rule that selects the most recent non-fogotten version of a fact (based on a timestamp and a fact_key). This rule is part of a Rust system with modules for geometry (vector dot product), hnsw (approximate nearest neighbor), and a journal (append-only log). The system also includes 'honest abstention': if the nearest knowledge is beyond a calibrated threshold (GEOLLM_THRESHOLD), it says 'I don't know'. In experiments on MQUAKE-CF (400 counterfactual cases), the six-line rule outperforms a pipeline using bge-m3 encoder and gemma3 arbiter. The author emphasizes that the rule yields four properties (versions, freshness, forgetting, historical view) for free, and abstention as a fifth.
- Abbreviations
- AGI = Artificial General Intelligence — общий искусственный интеллект
- RAG = Retrieval-Augmented Generation — генерация с опорой на поиск
- MQUAKE-CF = MQUAKE Counterfactual (dataset) — набор данных MQUAKE с контрфактическими фактами
Source: Habr — хаб ИИ —
original