\documentclass[12pt]{article} \usepackage[margin=1in]{geometry} \usepackage{cite} \usepackage{amsmath,amssymb,amsfonts} \usepackage{graphicx} \usepackage{booktabs} \usepackage{hyperref} \usepackage{setspace} \usepackage{caption} \usepackage{subcaption} \usepackage{float} \usepackage{xcolor} \usepackage{array} \usepackage{colortbl} \usepackage{microtype} \onehalfspacing \definecolor{wingreen}{rgb}{0.13, 0.55, 0.13} \definecolor{lightgray}{gray}{0.93} \definecolor{colA}{RGB}{44,123,182} \definecolor{colB}{RGB}{26,150,65} \definecolor{colC}{RGB}{244,109,67} \definecolor{colD}{RGB}{123,45,139} \hypersetup{colorlinks=true, linkcolor=black, citecolor=black, urlcolor=blue} \title{\textbf{\Large Scale vs.\ Architecture: A 2$\times$2 Empirical Study of\\ Model Size and Pipeline Design in Retrieval-Augmented Generation}} \author{Ahmad Tawil \\ \texttt{ahmadtawil.se@gmail.com}} \date{May 2026} \begin{document} \maketitle %────────────────────────────────────────────────────────────────────────────── \begin{abstract} %────────────────────────────────────────────────────────────────────────────── The prevailing assumption in Retrieval-Augmented Generation (RAG) is that larger models and more sophisticated pipelines always produce better results. This paper challenges that assumption with a controlled 2$\times$2 factorial experiment on the HotpotQA multi-hop question-answering benchmark, crossing two model families (GPT-5.4-mini and Llama-3.1-8B) with two pipeline architectures (naive single-pass RAG and agentic self-correcting RAG). The four systems---Goliath (Naive~GPT), David (Agentic~Llama), Hermes (Naive~Llama), and Titan (Agentic~GPT)---are evaluated on 100 questions across exact-match accuracy and four RAGAS metrics. The headline finding is a substantial \textbf{model-by-pipeline interaction effect}: agentic self-correction \emph{helps} GPT-5.4-mini ($+6.0$ percentage points in exact-match) but \emph{hurts} Llama-3.1-8B ($-7.0$pp), revealing that iterative self-correction imposes a reasoning demand that smaller models cannot reliably satisfy. Surprisingly, Hermes (Naive Llama) achieves the highest overall exact-match accuracy of all four systems (65.0\%), outperforming all GPT variants on this metric. On RAGAS metrics, Titan (Agentic GPT) leads in answer relevancy, context precision, and context recall, while Hermes leads in faithfulness. All four systems achieved 100\% completion rate, enabled by running inference on a local GPU (NVIDIA Tesla T4) via Ollama, eliminating the rate-limit failures that plagued earlier work. These results suggest that for smaller models, architectural complexity is a liability rather than an asset, and that the decision to deploy an agentic pipeline should be conditioned on the underlying model's self-correction capability. \end{abstract} \noindent\textbf{Keywords:} Retrieval-Augmented Generation, Agentic AI, Self-Correction, Multi-Hop Question Answering, HotpotQA, LangGraph, LLM Evaluation, RAGAS, Model Scale, Pipeline Architecture %────────────────────────────────────────────────────────────────────────────── \section{Introduction} %────────────────────────────────────────────────────────────────────────────── The scaling hypothesis---that more parameters yield better outputs---has dominated NLP research for the past decade~\cite{brown2020gpt3,kaplan2020scaling}. In the RAG setting, this translates to the common engineering practice of coupling the largest affordable frontier model with a simple retrieve-then-generate pipeline. A parallel line of work has shown that \emph{how} a model reasons may matter as much as \emph{how large} it is: agentic systems that iterate, self-check, and reformulate their queries achieve strong results on multi-step reasoning tasks~\cite{yao2022react,shinn2023reflexion,asai2023selfrag}. Prior work typically evaluates scale and architecture in isolation. In this paper I ask a more precise question: \emph{do scale and pipeline architecture interact?} Specifically, does an agentic pipeline provide the same benefit regardless of the underlying model's capability, or does it require a certain level of reasoning competence to function correctly? I answer this empirically with a \textbf{2$\times$2 factorial design}: \begin{center} \begin{tabular}{lcc} \toprule & \textbf{Naive Pipeline} & \textbf{Agentic Pipeline} \\ \midrule \textbf{GPT-5.4-mini} & System A (Goliath) & System D (Titan) \\ \textbf{Llama-3.1-8B} & System C (Hermes) & System B (David) \\ \bottomrule \end{tabular} \end{center} All four systems share identical embeddings, vector store, retrieval configuration, and benchmark, isolating model choice and pipeline architecture as the sole independent variables. My \textbf{contributions} are: \begin{enumerate} \item I conduct the first controlled 2$\times$2 factorial comparison of model scale and pipeline architecture in RAG, on the HotpotQA distractor benchmark. \item I observe an \textbf{interaction effect}: agentic self-correction benefits capable models but harms smaller ones, suggesting a \emph{capability threshold} for productive self-correction. \item I show that Naive Llama achieves the highest exact-match accuracy of all four systems, challenging the assumption that architectural sophistication always helps. \item I demonstrate 100\% completion across all systems by eliminating API rate limits through local GPU inference (NVIDIA Tesla T4 via Ollama), removing a confound present in earlier studies. \end{enumerate} %────────────────────────────────────────────────────────────────────────────── \section{Related Work} %────────────────────────────────────────────────────────────────────────────── \subsection{Retrieval-Augmented Generation} Lewis et al.~\cite{lewis2020retrieval} introduced the canonical RAG architecture, demonstrating that augmenting generation with retrieved passages substantially reduces hallucination on knowledge-intensive tasks. Subsequent work improved retrieval quality via dense passage retrieval~\cite{karpukhin2020dense} and hybrid sparse-dense strategies. Most practical deployments still follow the original single-pass retrieve-then-generate paradigm (our ``naive'' baseline). \subsection{Agentic and Iterative Reasoning} ReAct~\cite{yao2022react} showed that interleaving reasoning traces with action execution improves multi-step task performance. Reflexion~\cite{shinn2023reflexion} extended this with verbal self-reflection. Self-RAG~\cite{asai2023selfrag} introduced critic tokens that allow a model to decide when to retrieve, evaluate retrieved passages, and verify its own output---a conceptual precursor to the agentic pipeline implemented in this work. Critically, these prior works do not systematically study whether the benefit of agentic iteration depends on the capability of the underlying model. \subsection{Multi-Hop Question Answering} HotpotQA~\cite{yang2018hotpotqa} requires reasoning over two or more Wikipedia paragraphs. The distractor setting presents 10 paragraphs per question (2 gold, 8 irrelevant), specifically to test whether a system can identify relevant evidence rather than treating all retrieved text as equally useful. This makes HotpotQA especially sensitive to the document-grading component of agentic pipelines. \subsection{RAG Evaluation} Traditional n-gram metrics (BLEU, ROUGE) do not capture factual grounding quality. The RAGAS framework~\cite{es2023ragas} addresses this with four LLM-as-judge metrics: \emph{faithfulness}, \emph{answer relevancy}, \emph{context precision}, and \emph{context recall}. %────────────────────────────────────────────────────────────────────────────── \section{Methodology} %────────────────────────────────────────────────────────────────────────────── \subsection{System Architectures} \subsubsection{System A --- Goliath: Naive RAG with GPT-5.4-mini} The standard single-pass pipeline: embed question $\rightarrow$ retrieve top-5 passages from ChromaDB $\rightarrow$ generate answer with GPT-5.4-mini. No document grading, query reformulation, or post-generation verification. This represents the most common RAG implementation found in production today. \subsubsection{System B --- David: Agentic RAG with Llama-3.1-8B} A cyclic directed graph implemented with LangGraph. The same model (Llama-3.1-8B served via Ollama on a Tesla T4 GPU) performs all agentic reasoning steps: document grading, query rewriting, generation, and hallucination verification. The pipeline state tracks retrieval and generation attempt counts to enforce loop bounds. \textbf{Control flow:} After grading, if documents are relevant $\rightarrow$ generate; else if retrieval attempts $< 3$ $\rightarrow$ rewrite query and retry; else $\rightarrow$ generate from best available documents. After hallucination check, if answer is grounded $\rightarrow$ end; else if generation attempts $< 2$ $\rightarrow$ regenerate; else $\rightarrow$ end. \subsubsection{System C --- Hermes: Naive RAG with Llama-3.1-8B} Identical single-pass pipeline as System A but using Llama-3.1-8B (via Ollama) instead of GPT-5.4-mini. This cell isolates the effect of model choice under the naive pipeline. \subsubsection{System D --- Titan: Agentic RAG with GPT-5.4-mini} Identical agentic graph as System B but using GPT-5.4-mini (via OpenAI API) for all reasoning steps. This cell isolates the effect of pipeline architecture with the more capable model. \subsection{Shared Infrastructure} All four systems share identical infrastructure to ensure a fair comparison: \begin{itemize} \item \textbf{Embedding model:} \texttt{sentence-transformers/all-MiniLM-L6-v2}, running locally on CPU with L2-normalised embeddings. \item \textbf{Vector store:} ChromaDB (\texttt{hotpotqa\_dev} collection), with per-question corpus scoping via \texttt{question\_id} metadata. \item \textbf{Retrieval:} Top-$k=5$ passages per query, filtered to the current question's 10-paragraph partition. \item \textbf{Chunking:} Paragraphs stored as single chunks (RecursiveCharacterTextSplitter, max 1{,}000 chars, 200-char overlap). \item \textbf{Temperature:} 0.0 for all LLM calls (deterministic). \end{itemize} \subsection{Inference Infrastructure} Llama systems (B and C) run on a \textbf{NVIDIA Tesla T4 GPU} (15\,GiB VRAM) via Ollama on Google Colab, using the \texttt{llama3.1:8b} quantised model ($\approx$5.4\,GiB VRAM). GPT systems (A and D) call the OpenAI API. This setup achieved 100\% completion for all four systems, resolving the rate-limit failures reported in earlier work with cloud inference APIs. \subsection{Dataset} \begin{table}[H] \centering \caption{HotpotQA Benchmark Configuration} \label{tab:dataset} \begin{tabular}{ll} \toprule \textbf{Parameter} & \textbf{Value} \\ \midrule Split & Validation (distractor setting) \\ Sample size & 100 questions \\ Random seed & 42 \\ Difficulty & Hard (all sampled questions) \\ Question types & Bridge: 86 (86\%), Comparison: 14 (14\%) \\ Paragraphs ingested & 1{,}000 (100 $\times$ 10) \\ \bottomrule \end{tabular} \end{table} \subsection{Evaluation Metrics} \textbf{(1) Exact-Match Accuracy (EM):} The ground-truth answer string must appear as a substring of the system's response. A conservative metric that provides a direct, reproducible comparison signal. \textbf{(2) RAGAS Metrics:} Evaluated with GPT-5.4 as the judge LLM. Four metrics are computed for each system: \emph{faithfulness}, \emph{answer relevancy}, \emph{context precision}, and \emph{context recall}, each scored in $[0,1]$. All four metrics computed successfully for all systems, likely due to the updated RAGAS~0.4.x API and the slightly different 100-question sample relative to earlier work. %────────────────────────────────────────────────────────────────────────────── \section{Results} %────────────────────────────────────────────────────────────────────────────── \subsection{Completion Rate} All four systems completed all 100 questions (100\% completion rate). This is a significant improvement over earlier work, where Groq API rate limits caused a 30.5\% failure rate for the agentic Llama system. Running Llama locally on a Tesla T4 GPU via Ollama eliminates this constraint entirely. \subsection{Exact-Match Accuracy} \begin{table}[H] \centering \caption{Exact-Match Accuracy by System and Question Type} \label{tab:em} \rowcolors{2}{lightgray}{white} \begin{tabular}{lcccc} \toprule \textbf{System} & \textbf{Model} & \textbf{Pipeline} & \textbf{Overall EM} & \textbf{Bridge / Comparison} \\ \midrule A --- Goliath & GPT-5.4-mini & Naive & 52.0\% & 50.0\% / 64.3\% \\ B --- David & Llama-3.1-8B & Agentic & 58.0\% & 55.8\% / 71.4\% \\ \textcolor{wingreen}{\textbf{C --- Hermes}} & \textcolor{wingreen}{\textbf{Llama-3.1-8B}} & \textcolor{wingreen}{\textbf{Naive}} & \textcolor{wingreen}{\textbf{65.0\%}} & \textcolor{wingreen}{\textbf{64.0\%}} / \textcolor{wingreen}{\textbf{71.4\%}} \\ D --- Titan & GPT-5.4-mini & Agentic & 58.0\% & 55.8\% / 71.4\% \\ \bottomrule \end{tabular} \end{table} \begin{figure}[H] \centering \includegraphics[width=0.92\textwidth]{../evaluation/results/chart_exact_match.png} \caption{Exact-match accuracy by question type for all four systems. System C (Hermes --- Naive Llama) achieves the highest overall accuracy (65.0\%) across both bridge and comparison questions.} \label{fig:exactmatch} \end{figure} \subsection{RAGAS Evaluation} \begin{table}[H] \centering \caption{RAGAS Metric Scores --- All Four Systems} \label{tab:ragas} \rowcolors{2}{lightgray}{white} \begin{tabular}{lcccc} \toprule \textbf{Metric} & \textbf{A (Goliath)} & \textbf{B (David)} & \textbf{C (Hermes)} & \textbf{D (Titan)} \\ \midrule Faithfulness & 0.6025 & 0.7135 & \textcolor{wingreen}{\textbf{0.7333}} & 0.6431 \\ Answer Relevancy & 0.6595 & 0.5159 & 0.5405 & \textcolor{wingreen}{\textbf{0.6882}} \\ Context Precision & 0.6642 & 0.6882 & 0.6523 & \textcolor{wingreen}{\textbf{0.6896}} \\ Context Recall & 0.7500 & 0.7400 & 0.7600 & \textcolor{wingreen}{\textbf{0.7700}} \\ \midrule \textbf{Metrics Won}\textsuperscript{$\dagger$} & 0/4 & 0/4 & 1/4 & \textcolor{wingreen}{\textbf{3/4}} \\ \bottomrule \multicolumn{5}{l}{\footnotesize $\dagger$ Magnitudes of leads vary; this row indicates direction only.} \end{tabular} \end{table} \begin{figure}[H] \centering \includegraphics[width=\textwidth]{../evaluation/results/chart_ragas.png} \caption{RAGAS metric comparison across all four systems. System C (Hermes) leads on faithfulness; System D (Titan) leads on answer relevancy, context precision, and context recall.} \label{fig:ragas} \end{figure} \subsection{The 2$\times$2 Interaction Effect} The central finding of this study is that scale and architecture do not combine additively---they \emph{interact}. \begin{figure}[H] \centering \includegraphics[width=\textwidth]{../evaluation/results/chart_2x2_interaction.png} \caption{2$\times$2 interaction heatmaps. \textit{Left}: exact-match accuracy. \textit{Right}: mean RAGAS score. Arrows show the effect of switching from naive to agentic pipeline within each model family. The agentic pipeline \emph{helps} GPT but \emph{hurts} Llama on EM; results are more nuanced on RAGAS.} \label{fig:interaction} \end{figure} Table~\ref{tab:interaction} quantifies the main effects and interaction: \begin{table}[H] \centering \caption{Main Effects and Interaction (Exact-Match Accuracy)} \label{tab:interaction} \begin{tabular}{lcc} \toprule \textbf{Effect} & \textbf{Naive} & \textbf{Agentic} \\ \midrule GPT-5.4-mini row & 52.0\% (A) & 58.0\% (D) \quad $\Delta = +6.0$pp \\ Llama-3.1-8B row & \textbf{65.0\%} (C) & 58.0\% (B) \quad $\Delta = -7.0$pp \\ \midrule \multicolumn{3}{l}{\textbf{Interaction}: $+6.0 - (-7.0) = 13.0$pp \quad (agentic benefits GPT more than Llama)} \\ \bottomrule \end{tabular} \end{table} The 13-percentage-point interaction term is the key result: the agentic pipeline is far from model-agnostic. \subsection{Summary of Results} \begin{table}[H] \centering \caption{Full Metric Summary} \label{tab:summary} \begin{tabular}{lcccc} \toprule \textbf{Metric} & \textbf{A (Goliath)} & \textbf{B (David)} & \textbf{C (Hermes)} & \textbf{D (Titan)} \\ \midrule EM Overall & 52.0\% & 58.0\% & \textcolor{wingreen}{\textbf{65.0\%}} & 58.0\% \\ EM Bridge & 50.0\% & 55.8\% & \textcolor{wingreen}{\textbf{64.0\%}} & 55.8\% \\ EM Comparison & 64.3\% & 71.4\% & \textcolor{wingreen}{\textbf{71.4\%}} & \textcolor{wingreen}{\textbf{71.4\%}} \\ RAGAS Faithfulness & 0.603 & 0.714 & \textcolor{wingreen}{\textbf{0.733}} & 0.643 \\ RAGAS Ans.\ Rel. & 0.660 & 0.516 & 0.541 & \textcolor{wingreen}{\textbf{0.688}} \\ RAGAS Ctx.\ Prec. & 0.664 & 0.688 & 0.652 & \textcolor{wingreen}{\textbf{0.690}} \\ RAGAS Ctx.\ Rec. & 0.750 & 0.740 & 0.760 & \textcolor{wingreen}{\textbf{0.770}} \\ Completion Rate & \textcolor{wingreen}{\textbf{100\%}} & \textcolor{wingreen}{\textbf{100\%}} & \textcolor{wingreen}{\textbf{100\%}} & \textcolor{wingreen}{\textbf{100\%}} \\ \bottomrule \end{tabular} \end{table} \subsection{Agentic Loop Behaviour} \label{sec:loop} \begin{table}[H] \centering \caption{Agentic Loop Statistics ($n = 100$ questions each)} \label{tab:loop} \rowcolors{2}{lightgray}{white} \begin{tabular}{lcc} \toprule \textbf{Statistic} & \textbf{System B (David)} & \textbf{System D (Titan)} \\ \midrule Mean retrieval attempts & 2.78 & 2.76 \\ Mean generation attempts & 1.07 & 1.31 \\ \bottomrule \end{tabular} \end{table} \begin{figure}[H] \centering \includegraphics[width=\textwidth]{../evaluation/results/chart_loop_behaviour.png} \caption{Agentic loop behaviour for System B (Llama) and System D (GPT) across 100 questions. Both systems exhaust the retrieval ceiling at similar rates. System D triggers significantly more regeneration passes (mean 1.31 vs 1.07), reflecting GPT's stricter hallucination detection.} \label{fig:loop} \end{figure} Both agentic systems exhaust the retrieval budget at similar rates (2.78 vs 2.76 mean attempts), indicating that the retrieval bottleneck is driven by the distractor corpus structure rather than model capability. However, System D (GPT) triggers substantially more regeneration passes than System B (Llama): mean 1.31 vs 1.07. This suggests that GPT's hallucination checker is stricter and more accurate---it correctly identifies unsupported claims and demands regeneration, whereas Llama's checker is more permissive and often accepts a flawed first draft. %────────────────────────────────────────────────────────────────────────────── \section{Discussion} %────────────────────────────────────────────────────────────────────────────── \subsection{Evidence for a Capability Threshold in Self-Correction} The most important finding of this paper is the model-by-pipeline interaction. The agentic pipeline imposes several complex reasoning demands beyond simple generation: it requires the model to \emph{grade} retrieved passages for relevance, \emph{rewrite} queries when retrieval fails, and \emph{verify} factual grounding in its own output. These tasks require meta-cognitive ability that may exceed the capacity of smaller models. Llama-3.1-8B appears to fall below a \textbf{capability threshold} for productive self-correction. Evidence for this comes from three observations: (1) the agentic pipeline \emph{hurts} Llama on EM ($-7.0$pp), while it helps GPT ($+6.0$pp); (2) Llama's hallucination checker is permissive, accepting first drafts at a high rate (mean 1.07 generation attempts vs 1.31 for GPT), suggesting it does not reliably detect unsupported claims; and (3) Naive Llama (65.0\% EM) outperforms Agentic Llama (58.0\% EM), meaning the self-correction loop introduces more errors than it corrects. \subsection{Why Naive Llama Wins on Exact-Match} The result that Hermes (Naive Llama, 65.0\% EM) outperforms all GPT variants on exact-match accuracy is counterintuitive. We attribute this to two factors. First, Llama-3.1-8B's responses tend to be more concise and direct, making exact-match---a substring check---more likely to succeed. GPT-5.4-mini often produces fluent but verbose answers that restate the question or add context, causing the short ground-truth string to be absent even when the answer is semantically correct. Second, the distractor corpus structure may favour Llama's pattern-matching behaviour on HotpotQA's factoid question style. This finding highlights a known limitation of exact-match as an evaluation metric: it penalises verbosity even when the underlying answer is correct. The RAGAS metrics, which evaluate semantic quality rather than string overlap, tell a different story: System D (Agentic GPT) leads on three of four RAGAS metrics, suggesting it produces higher-quality answers that happen to be phrased differently from the ground truth. \subsection{Faithfulness vs.\ Answer Relevancy Trade-off} An interesting split emerges on RAGAS metrics: Hermes (Naive Llama) leads on \emph{faithfulness} (0.733), while Titan (Agentic GPT) leads on \emph{answer relevancy} (0.688). Faithfulness measures whether every claim in the answer is grounded in retrieved context; answer relevancy measures whether the answer actually addresses the question. Llama's higher faithfulness may reflect a tendency to stay closely anchored to the retrieved text (avoiding hallucination by paraphrasing directly), while GPT's higher answer relevancy reflects its stronger instruction-following and question-interpretation ability. The agentic pipeline's hallucination checker further reinforces faithfulness in both systems. \subsection{Implications for Practitioners} \begin{itemize} \item \textbf{Match pipeline complexity to model capability.} An agentic self-correction loop is not universally beneficial. It requires the model to perform document grading, query rewriting, and hallucination verification reliably. For smaller models such as Llama-3.1-8B, a well-tuned naive pipeline may outperform an agentic one. \item \textbf{Evaluate metrics carefully.} Exact-match favours concise models; RAGAS metrics favour semantically accurate but potentially verbose ones. Use multiple complementary metrics before concluding which system is ``better.'' \item \textbf{Local GPU inference eliminates rate limits.} Running Llama on a Tesla T4 via Ollama achieved 100\% completion at effectively zero marginal cost, compared to the 30.5\% failure rate observed with cloud API inference. This is a practical recommendation for any project that issues high volumes of LLM calls. \item \textbf{Naive RAG remains competitive.} Despite the attention agentic systems receive, a simple retrieve-then-generate pipeline with a capable model or a concise smaller model remains a strong baseline that should not be dismissed. \end{itemize} \subsection{Limitations} \begin{itemize} \item \textbf{Single benchmark.} Results are specific to HotpotQA multi-hop distractor QA. The interaction effect may differ on single-hop, conversational, or domain-specific benchmarks. \item \textbf{Exact-match conservatism.} The substring EM metric under-credits verbose but correct answers, systematically disadvantaging GPT systems. \item \textbf{Single experimental run.} No confidence intervals are reported. Despite temperature 0, minor variance may exist across runs. \item \textbf{Quantised local model.} The Llama model served via Ollama uses 4-bit quantisation, which may introduce small accuracy penalties relative to the full-precision model used via cloud APIs. \item \textbf{Single model pair.} The capability threshold hypothesis requires further validation with additional model sizes (e.g., Llama-3.1-70B, Llama-3.2-3B) to establish where the threshold lies. \item \textbf{Small comparison-question subset.} Only 14 of the 100 sampled questions are of the comparison type; results broken down by question type for this subset should be read as indicative rather than definitive. \end{itemize} %────────────────────────────────────────────────────────────────────────────── \section{Future Work} %────────────────────────────────────────────────────────────────────────────── \begin{itemize} \item \textbf{Capability threshold study.} Systematically varying model size (1B, 3B, 8B, 13B, 70B parameters) within the same agentic pipeline to identify the parameter count at which self-correction becomes productive. \item \textbf{Full-scale evaluation.} Running on the full HotpotQA validation set (7{,}405 questions) for statistically robust results. \item \textbf{Comparison with other agentic frameworks.} Contrasting LangGraph self-correction with ReAct~\cite{yao2022react}, Self-RAG~\cite{asai2023selfrag}, and Reflexion~\cite{shinn2023reflexion}. \item \textbf{Semantic evaluation.} Replacing or supplementing exact-match with model-based semantic equivalence scoring to reduce the verbosity penalty on GPT systems. \item \textbf{Cost-performance analysis.} Measuring API cost and inference latency per correct answer across all four systems to compute a cost-adjusted performance frontier. \end{itemize} %────────────────────────────────────────────────────────────────────────────── \section{Conclusion} %────────────────────────────────────────────────────────────────────────────── This paper presented a controlled 2$\times$2 factorial study crossing two model families (GPT-5.4-mini, Llama-3.1-8B) with two pipeline architectures (naive RAG, agentic self-correcting RAG) on the HotpotQA distractor benchmark. The central finding is a \textbf{13-percentage-point interaction effect} in exact-match accuracy: agentic self-correction improves GPT-5.4-mini by 6 percentage points but degrades Llama-3.1-8B by 7 percentage points. This result challenges the assumption that agentic pipelines are universally beneficial, suggesting instead that iterative self-correction requires a \emph{capability threshold} in the underlying model that Llama-3.1-8B does not reliably meet. Practically, Hermes (Naive Llama) achieves the highest overall exact-match accuracy (65.0\%), demonstrating that a simpler pipeline with a well-matched model can outperform more complex alternatives. On RAGAS metrics, the picture is more nuanced: Titan (Agentic GPT) leads on answer relevancy, context precision, and context recall, while Hermes leads on faithfulness---suggesting a faithfulness-vs-relevancy trade-off between the two model families. Finally, by switching from cloud API inference to local GPU inference via Ollama on a Tesla T4, we achieved 100\% completion across all four systems, eliminating the rate-limit failures that have confounded earlier comparisons of agentic RAG systems. Our results suggest that practitioners and researchers should choose an agentic pipeline based on their model's self-correction capability, rather than assuming that more complexity is always better. %────────────────────────────────────────────────────────────────────────────── \begin{thebibliography}{00} %────────────────────────────────────────────────────────────────────────────── \bibitem{lewis2020retrieval} P.\ Lewis et al., ``Retrieval-augmented generation for knowledge-intensive NLP tasks,'' \textit{Advances in Neural Information Processing Systems}, vol.~33, pp.~9459--9474, 2020. \bibitem{karpukhin2020dense} V.\ Karpukhin et al., ``Dense passage retrieval for open-domain question answering,'' in \textit{Proc.\ EMNLP}, 2020, pp.~6769--6781. \bibitem{yang2018hotpotqa} Z.\ Yang et al., ``HotpotQA: A dataset for diverse, explainable multi-hop question answering,'' in \textit{Proc.\ EMNLP}, 2018, pp.~2369--2380. \bibitem{brown2020gpt3} T.\ Brown et al., ``Language models are few-shot learners,'' \textit{Advances in Neural Information Processing Systems}, vol.~33, pp.~1877--1901, 2020. \bibitem{kaplan2020scaling} J.\ Kaplan et al., ``Scaling laws for neural language models,'' \textit{arXiv preprint arXiv:2001.08361}, 2020. \bibitem{yao2022react} S.\ Yao et al., ``ReAct: Synergizing reasoning and acting in language models,'' in \textit{Proc.\ ICLR}, 2023. \bibitem{shinn2023reflexion} N.\ Shinn et al., ``Reflexion: Language agents with verbal reinforcement learning,'' \textit{Advances in Neural Information Processing Systems}, vol.~36, 2023. \bibitem{asai2023selfrag} A.\ Asai et al., ``Self-RAG: Learning to retrieve, generate, and critique through self-reflection,'' in \textit{Proc.\ ICLR}, 2024. \bibitem{es2023ragas} S.\ Es et al., ``RAGAS: Automated evaluation of retrieval augmented generation,'' \textit{arXiv preprint arXiv:2309.15217}, 2023. \end{thebibliography} \end{document}