M\'as contexto no es mejor. Paradoja de la diluci\'on vectorial en RAG corporativos

TL;DR

This paper investigates the impact of summary injection ratios in RAG systems, revealing a trade-off where moderate levels improve recall but excessive injection diminishes precision, and proposes a theoretical method to optimize this ratio.

Contribution

It introduces a theoretical framework to determine the optimal injection ratio in RAG, balancing recall and precision by analyzing the vector dilution effect.

Findings

Moderate injection increases recall by 18%.

Exceeding CIR > 0.4 reduces precision by 22%.

Proposes a theoretical model for optimal injection ratio.

Abstract

T\'ecnicas recientes de "Contextualized Chunking" inyectan res\'umenes para mejorar el contexto en RAG, pero introducen una "diluci\'on vectorial" que opaca el contenido local. Evaluando distintos ratios de inyecci\'on, demostramos una curva en "U invertida": una inyecci\'on moderada mejora el "Recall" (+18%), pero superar un umbral cr\'itico (CIR > 0.4) reduce la precisi\'on en un 22% para consultas espec\'ificas. Proponemos un marco te\'orico para calcular el ratio \'optimo de inyecci\'on. -- Recent "Contextualized Chunking" techniques inject summaries to improve RAG context but introduce "vector dilution" drowning out local content. Evaluating various injection ratios, we demonstrate an "inverted U" curve: moderate injection boosts Recall (+18%), but exceeding a critical threshold (CIR > 0.4) drops precision by 22% for specific queries. We propose a theoretical framework to calculate the optimal injection ratio.