Scale vs. Architecture: A 2×2 Empirical Study of RAG

Does an agentic self-correction pipeline always beat a simpler one? Does a bigger model? This project answers both questions with a controlled 2×2 factorial experiment.

$Read Paper$

Experimental Design #

Four systems cross two model families with two pipeline architectures:

	Naive Pipeline	Agentic Pipeline
GPT-5.4-mini	System A — Goliath	System D — Titan
Llama-3.1-8B	System C — Hermes	System B — David

All four systems share identical embeddings, vector store, retrieval configuration, and benchmark — isolating model choice and pipeline architecture as the sole independent variables.

Results #

Evaluated on 100 questions from the HotpotQA distractor benchmark (seed = 42):

Exact-Match Accuracy #

System	Model	Pipeline	Overall	Bridge	Comparison
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%
C — Hermes	Llama-3.1-8B	Naive	65.0%	64.0%	71.4%
D — Titan	GPT-5.4-mini	Agentic	58.0%	55.8%	71.4%

RAGAS Metrics (judge: GPT-5.4) #

Metric	A — Goliath	B — David	C — Hermes	D — Titan
Faithfulness	0.603	0.714	0.733	0.643
Answer Relevancy	0.660	0.516	0.541	0.688
Context Precision	0.664	0.688	0.652	0.690
Context Recall	0.750	0.740	0.760	0.770

The Key Finding — Interaction Effect #

The agentic pipeline is not model-agnostic:

	Naive → Agentic
GPT-5.4-mini	+6.0 pp
Llama-3.1-8B	−7.0 pp

13 pp interaction term. Agentic self-correction helps the more capable model but hurts the smaller one — suggesting a capability threshold for productive self-correction. All four systems achieved 100% completion by running Llama on a local GPU (NVIDIA Tesla T4 via Ollama), eliminating the rate-limit failures of earlier work.

Charts #

Exact-Match Accuracy by Question Type #

Exact Match

RAGAS Evaluation — All Four Systems #

RAGAS

2×2 Interaction: Model Scale vs Pipeline Architecture #

2x2 Interaction

Agentic Loop Behaviour — Systems B and D #

Loop Behaviour

Systems #

	A — Goliath	B — David	C — Hermes	D — Titan
Model	GPT-5.4-mini	Llama-3.1-8B	Llama-3.1-8B	GPT-5.4-mini
Pipeline	Naive RAG	Agentic RAG	Naive RAG	Agentic RAG
Orchestration	LangChain	LangGraph	LangChain	LangGraph
Self-correction	None	Doc grading + query rewriting + hallucination check	None	Doc grading + query rewriting + hallucination check

Agentic Pipeline (Systems B and D) #

Question
   │
   ▼
Retrieve (ChromaDB, top-5)
   │
   ▼
Grade Documents ──── < 50% relevant? ──── Rewrite Query ──┐
   │                                                       │
   │ ≥ 50% relevant (or max 3 retries)               (loop back)
   ▼
Generate Answer
   │
   ▼
Check Hallucination ──── not grounded? ──── (regenerate, max 2×)
   │
   ▼
Answer

Benchmark #

HotpotQA — distractor setting, validation split

Parameter	Value
Questions sampled	100 (seed = 42)
Paragraphs ingested	1,000 (100 × 10)
Gold paragraphs per question	2
Distractor paragraphs per question	8
Question types	Bridge (86), Comparison (14)
Difficulty	Hard (all)

Project Structure #

agentic-rag-vs-scale/
├── config.py                        # Central configuration (all models, limits, paths)
├── requirements.txt                 # Dependencies
├── .env.example                     # API key template
├── check_apis.py                    # Test OpenAI, Groq, and Ollama backends
│
├── ingestion/
│   └── ingest_hotpotqa.py           # Download HotpotQA + build ChromaDB corpus
│
├── systems/
│   ├── llm_factory.py               # LLM backend selector (Groq / Ollama / local GPU)
│   ├── system_a_goliath.py          # Naive RAG + GPT-5.4-mini
│   ├── system_b_david.py            # Agentic RAG + Llama-3.1-8B
│   ├── system_c_hermes.py           # Naive RAG + Llama-3.1-8B
│   └── system_d_titan.py            # Agentic RAG + GPT-5.4-mini
│
├── evaluation/
│   ├── load_benchmark.py            # Load sampled questions
│   ├── run_tournament.py            # Run all four systems on all questions
│   ├── evaluate_ragas.py            # Score with RAGAS 0.4.x framework
│   ├── visualize_results.py         # Generate publication charts
│   └── results/                     # Output JSON, CSV, and PNG files
│
└── agentic-rag-arxiv/
    └── main.tex                     # LaTeX research paper

Quick Start #

1. Clone and install #

git clone https://github.com/AhmadTawil1/agentic-rag-vs-scale.git
cd agentic-rag-vs-scale
pip install -r requirements.txt

2. Set API keys #

copy .env.example .env

Edit .env:

OPENAI_API_KEY=your_openai_key
GROQ_API_KEY=your_groq_key          # for Llama via Groq (optional)
GROQ_API_KEY_2=your_backup_groq_key # backup key (optional)

To run Llama locally via Ollama instead of Groq (recommended — no rate limits):

USE_OLLAMA=true

3. Build the corpus #

Downloads HotpotQA from HuggingFace and ingests 1,000 paragraphs into ChromaDB.

python ingestion/ingest_hotpotqa.py

4. Run the tournament #

python evaluation/run_tournament.py

All four systems answer all 100 questions. Results are saved to evaluation/results/.

Llama rate limits: Systems B and C each make multiple LLM calls per question. Running on Groq's free tier will hit daily token limits. Use USE_OLLAMA=true with a local GPU (≥8 GB VRAM) to eliminate this constraint entirely.

5. Evaluate with RAGAS #

Run locally after downloading the result JSON files:

python evaluation/evaluate_ragas.py \
  --system-a evaluation/results/system_a_results_<TIMESTAMP>.json \
  --system-b evaluation/results/system_b_results_<TIMESTAMP>.json \
  --system-c evaluation/results/system_c_results_<TIMESTAMP>.json \
  --system-d evaluation/results/system_d_results_<TIMESTAMP>.json

6. Generate charts #

python evaluation/visualize_results.py \
  --system-a evaluation/results/system_a_results_<TIMESTAMP>.json \
  --system-b evaluation/results/system_b_results_<TIMESTAMP>.json \
  --system-c evaluation/results/system_c_results_<TIMESTAMP>.json \
  --system-d evaluation/results/system_d_results_<TIMESTAMP>.json \
  --comparison-csv evaluation/results/comparison.csv \
  --output-dir evaluation/results

Configuration #

Key parameters in config.py:

Parameter	Default	Description
`HOTPOTQA_SAMPLE_SIZE`	`100`	Questions sampled from validation split
`HOTPOTQA_SEED`	`42`	Random seed for reproducibility
`RETRIEVAL_TOP_K`	`5`	Documents retrieved per query
`MAX_RETRIEVAL_LOOPS`	`3`	Max query rewrites (agentic systems)
`MAX_GENERATION_RETRIES`	`2`	Max regeneration attempts (agentic systems)
`GOLIATH_MODEL`	`gpt-5.4-mini`	System A model
`TITAN_MODEL`	`gpt-5.4-mini`	System D model
`DAVID_MODEL`	`llama-3.1-8b-instant`	System B model (via Groq)
`HERMES_MODEL`	`llama-3.1-8b-instant`	System C model (via Groq)
`EVALUATOR_MODEL`	`gpt-5.4`	RAGAS judge model
`EMBEDDING_MODEL`	`all-MiniLM-L6-v2`	Shared embedding model (local)

Paper #

The full research paper is in agentic-rag-arxiv/main.tex.

To compile (requires MiKTeX or Overleaf):

cd agentic-rag-arxiv
pdflatex main.tex
pdflatex main.tex   # second pass for cross-references

Author #

Ahmad Tawil — ahmadtawil.se@gmail.com