quco-rag

name	quco-rag
title	QuCo-RAG: Quantifying Uncertainty for Dynamic RAG
version	0.0.2
engine	skillxiv-v0.0.2-claude-opus-4.6
license	MIT
url	https://arxiv.org/abs/2512.19134
keywords	["rag","retrieval","uncertainty","hallucination","calibration"]
description	Replace unreliable model-internal confidence signals with objective corpus statistics to decide when RAG retrieval is necessary. Pre-evaluates entity rarity in training data and verifies entity co-occurrence at runtime, triggering retrieval only when hallucination risk is high—improving reliability without per-model tuning.

Overview

QuCo-RAG addresses a critical flaw in dynamic RAG systems: relying on LLM confidence scores for retrieval decisions when models are notoriously poorly calibrated. This framework shifts to objective, corpus-based evidence that's reliably indicative of hallucination risk.

Core Technique

The key innovation is grounding retrieval decisions in pre-training corpus statistics rather than model outputs.

Pre-Generation Entity Assessment: Before generation, identify knowledge gaps by checking entity rarity in training data.

# Entity rarity assessment using corpus statistics
import infini_gram  # Trillion-token corpus query tool

class CorpusBasedUncertainty:
    def __init__(self, corpus_client):
        self.corpus = corpus_client  # Infini-gram for fast queries

    def assess_pre_generation(self, input_question):
        """
        Identify entities in question appearing rarely in training corpus.
        High rarity → high hallucination risk → retrieve.
        """
        entities = extract_entities(input_question)
        high_risk_entities = []

        for entity in entities:
            # Query corpus for entity frequency
            frequency = self.corpus.query_frequency(entity)

            # Low frequency = long-tail knowledge gap
            if frequency < RARITY_THRESHOLD:
                high_risk_entities.append(entity)
                return True  # Trigger retrieval

        return False  # Confident in internal knowledge

Runtime Verification via Co-occurrence: After generation, verify factual claims by checking entity co-occurrence in corpus.

def ():
    
    claims = extract_factual_claims(generated_text)
    hallucination_risk = 

     claim  claims:
        subject, predicate, obj = parse_claim(claim)

        
        cooccurrence = corpus_client.query_cooccurrence(
            entities=[subject, obj],
            context=predicate
        )

         cooccurrence == :
            
            hallucination_risk = 
            

     hallucination_risk