RAG 4: when to test more complex techniques
When I would test late chunking, visual retrieval, hierarchies or agents based on the errors that appear in search.
In this article
Recent literature does not offer one winning architecture for every document collection. It offers interventions addressing different problems. My job is to connect each intervention to an observed failure and an experiment capable of disproving its usefulness. Being newer does not automatically make a technique the next step for Taesa.
When a chunk loses context
Late Chunking contextualizes tokens from a long text before pooling their representations into chunks. It differs from generating a summary and from increasing overlap. It requires a model and interface supporting that processing. I would test it when cross-section references hurt retrieval, comparing it with a much simpler extracted prefix.
Late Chunking: method and experiments
When the evidence is visual
ColPali retrieves pages using images and multi-vector representations. It is relevant when extracted text loses tables or layout relationships. But finding the page does not prove that a model reads a number correctly. My evaluation would separate page retrieval, table interpretation and final answering, while measuring storage and latency.
ColPali and the ViDoRe benchmark
When an answer depends on several pieces
RAPTOR proposes a summary hierarchy for retrieval at different scales. I would consider it for global questions while retaining links to original passages. An explicit graph may help when relationships between contracts, entities and acts matter. Neither approach turns inferred relationships into verified facts. First I would build a small, reviewable case that conventional search cannot solve.
RAPTOR: hierarchical retrieval
Search agents add decisions and calls: they can discover a second query from the first piece of evidence. This also introduces planning failures, loops and cost. HERB studies heterogeneous enterprise sources in a synthetic software and product environment. Its difficulties justify multi-document tests, not a promise that adding agents will solve regulatory retrieval.
HERB: search over heterogeneous enterprise data
What a 2026 financial comparison supports
From BM25 to Corrective RAG found a strong combination of hybrid retrieval and reranking, with BM25 outperforming the studied dense retriever. However, it retrieves whole documents, uses one main embedding model and evaluates numerical answers. It does not establish the best chunking or regulatory system. I would use it to choose baselines, not copy a universal ranking.
2026 financial benchmark and its limitations
My sequence would be to resolve identity and extraction, compare simple retrievers, and only then add complexity targeting a failure. Recent works consulted on arXiv are research evidence; without additional publication verification I do not present them as peer-reviewed consensus. The decisive test remains a frozen local evaluation with explicit cost and limitations.
D-RAC, also from September 2026, separates multimodal Markdown conversion from chunk planning over identifiers. That separation may reduce repeated segmentation costs. However, a run without conversion or chunking errors does not establish semantic correctness or better final answers. I would preserve tables and originals to check whether prose conversion retains units and relationships; I would not convert every format to PDF without comparing information loss.
D-RAC: conversion and segmentation over identifiers