Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Bidirectional RAG: Safe Self-Improving Retrieval-Augmented Generation Through Multi-Stage Validation

Teja Chinthala

TL;DR

The paper addresses the limitation of static corpora in retrieval-augmented generation (RAG) systems and explores safe self-improvement by enabling validated write-back of model outputs into the retrieval corpus. It introduces Bidirectional RAG, which adds a backward path governed by a multi-stage acceptance layer that combines grounding verification, attribution checking, and novelty detection to prevent hallucination pollution while allowing knowledge growth. An experience store captures critique signals from accepted and rejected outputs for meta-learning. Across four datasets, Bidirectional RAG achieves approximately a 2x gain in coverage while adding roughly 72% fewer documents than naive write-back, demonstrating feasible safe corpus expansion under explicit safety constraints such as $H(D_t) <= ext{epsilon}_h$ and $alpha(D_t) <= alpha_{ ext{max}}$. The work provides a practical pathway for deploying RAG systems that learn from deployment without retraining, balancing performance, safety, and computational overhead.

Abstract

Retrieval-Augmented Generation RAG systems enhance large language models by grounding responses in external knowledge bases, but conventional RAG architectures operate with static corpora that cannot evolve from user interactions. We introduce Bidirectional RAG, a novel RAG architecture that enables safe corpus expansion through validated write back of high quality generated responses. Our system employs a multi stage acceptance layer combining grounding verification (NLI based entailment, attribution checking, and novelty detection to prevent hallucination pollution while enabling knowledge accumulation. Across four datasets Natural Questions, TriviaQA, HotpotQA, Stack Overflow with three random seeds 12 experiments per system, Bidirectional RAG achieves 40.58% average coverage nearly doubling Standard RAG 20.33% while adding 72% fewer documents than naive write back 140 vs 500. Our work demonstrates that self improving RAG is feasible and safe when governed by rigorous validation, offering a practical path toward RAG systems that learn from deployment.

Bidirectional RAG: Safe Self-Improving Retrieval-Augmented Generation Through Multi-Stage Validation

TL;DR

The paper addresses the limitation of static corpora in retrieval-augmented generation (RAG) systems and explores safe self-improvement by enabling validated write-back of model outputs into the retrieval corpus. It introduces Bidirectional RAG, which adds a backward path governed by a multi-stage acceptance layer that combines grounding verification, attribution checking, and novelty detection to prevent hallucination pollution while allowing knowledge growth. An experience store captures critique signals from accepted and rejected outputs for meta-learning. Across four datasets, Bidirectional RAG achieves approximately a 2x gain in coverage while adding roughly 72% fewer documents than naive write-back, demonstrating feasible safe corpus expansion under explicit safety constraints such as and . The work provides a practical pathway for deploying RAG systems that learn from deployment without retraining, balancing performance, safety, and computational overhead.

Abstract

Retrieval-Augmented Generation RAG systems enhance large language models by grounding responses in external knowledge bases, but conventional RAG architectures operate with static corpora that cannot evolve from user interactions. We introduce Bidirectional RAG, a novel RAG architecture that enables safe corpus expansion through validated write back of high quality generated responses. Our system employs a multi stage acceptance layer combining grounding verification (NLI based entailment, attribution checking, and novelty detection to prevent hallucination pollution while enabling knowledge accumulation. Across four datasets Natural Questions, TriviaQA, HotpotQA, Stack Overflow with three random seeds 12 experiments per system, Bidirectional RAG achieves 40.58% average coverage nearly doubling Standard RAG 20.33% while adding 72% fewer documents than naive write back 140 vs 500. Our work demonstrates that self improving RAG is feasible and safe when governed by rigorous validation, offering a practical path toward RAG systems that learn from deployment.
Paper Structure (28 sections, 6 equations, 2 figures, 2 tables)

This paper contains 28 sections, 6 equations, 2 figures, 2 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Retrieval-Augmented Generation
  4. Hallucination Prevention
  5. Continual Learning
  6. Problem Formulation
  7. Approach
  8. System Architecture
  9. Multi-Stage Acceptance Layer
  10. Grounding Verification
  11. Attribution Checking
  12. Novelty Detection
  13. Experience Store
  14. Experimental Setup
  15. Datasets
  16. ...and 13 more sections

Figures (2)

  • Figure 1: Coverage and corpus growth comparison. Bidirectional RAG achieves substantial coverage gains while adding 72% fewer documents than Naive Write-back.
  • Figure 2: Coverage by dataset showing domain-specific variance. Stack Overflow benefits from strong initial corpus alignment, while sparse domains (NQ, HotpotQA) show more modest but consistent gains. The high standard deviations reflect genuine domain variance rather than experimental noise.