A-RAG: Scaling Agentic Retrieval via Hierarchical Interfaces скачать в хорошем качестве

A-RAG: Scaling Agentic Retrieval via Hierarchical Interfaces

Paper: https://arxiv.org/pdf/2602.03442 Notes: Problem — static RAG pipelines or fixed workflows; model cannot steer retrieval decisions. Solution — A-RAG: expose hierarchical retrieval interfaces so LLM can autonomously choose strategy and iterate. HIERARCHICAL INDEX: chunk corpus ≈1,000 tokens; decompose chunks into sentences; compute sentence-level embeddings mapped to parent chunk. TOOLS: keyword_search (exact-match scoring by frequency×keyword-length, returns snippets), semantic_search (sentence embeddings, chunk score = max sentence similarity), chunk_read (full chunk ± adjacent). AGENT LOOP: ReAct-like single-tool-per-iteration loop (Algorithm 1); LLM decides next tool or final answer; stops on answer or max iterations. CONTEXT TRACKER: maintain C_read set; repeated chunk_read returns “already read” with zero token cost to avoid redundancy. PROGRESSIVE ACQUISITION HEURISTIC: search → inspect snippets → selective chunk_read; minimizes retrieved tokens vs naive agentic retrieval. TRAINING / SIGNAL: no special SFT/RL required for framework; relies on prompt + tool descriptions; evaluated with GPT-4o-mini and GPT-5-mini backbones. INFERENCE LOGIC: agent picks best-granularity tool per step (keyword for precise entities, semantic for concept matches, chunk_read for full evidence); can read adjacents when context incomplete. ABLATION INSIGHT: removing keyword or semantic search or chunk_read hurts accuracy; full hierarchical set yields best results (ablation Table 2). SCALING BEHAVIOR: more max steps and higher reasoning effort improve accuracy; stronger models (GPT-5-mini) benefit more (e.g., 5→20 steps ≈ +8% for GPT-5-mini vs ≈+4% for GPT-4o-mini). FAILURE MODES SHIFT: bottleneck moves from “cannot find docs” to “reasoning-chain errors”; dominant secondary errors = entity confusion, wrong strategy, question misunderstanding. Source: A-RAG — "A-RAG: Scaling Agentic Retrieval-Augmented Generation via Hierarchical Retrieval Interfaces." Disclaimer: This is an AI-powered production. The scripts, insights, and voices featured in this podcast are generated entirely by Artificial Intelligence models. While we strive for technical accuracy by grounding our episodes in original research papers, listeners are encouraged to consult the primary sources for critical applications.