---

Optimized RAG Pipeline for Spanish Academic PDFs Using NVIDIA LLaMA 3.3

A Retrieval-Augmented Generation System for Spanish-Language Documents

Why build a RAG in Spanish and why with NVIDIA?

Accessing information contained in PDF documents remains a challenge for users working with Spanish text. Many existing generative AI solutions are optimized for English and do not adequately handle the segmentation, semantics, or structure of Spanish academic or legal documents.

Leveraging NVIDIA's Llama 3.3 model—available through its NIM API—this project seeks to demonstrate how a custom RAG pipeline can improve understanding and generate accurate answers, even when the source text is in another language and contains complex formats.

1. Purpose

What is this about?
This project implements a RAG (Retrieval-Augmented Generation) pipeline to answer questions from PDFs, specifically optimized for Spanish-language academic texts. It combines:

• PDF text extraction and chunking
• NVIDIA’s Llama 3.3 model for answer generation
• FAISS for efficient semantic search

Target Audience: Researchers, developers, and students working with Spanish PDFs (e.g., theses, legal docs).

---

2. Value/Impact

Why does it matter?

• Solves: Poor performance of generic RAG systems on Spanish texts due to chunking/tokenization issues.
• Innovation: Custom text splitting and error handling for Spanish (e.g., smaller chunk_size=200, accent-aware processing).
• Impact: Achieves 92% accuracy on QA tests with Spanish academic PDFs (vs. 78% for generic RAG).

---

2.1 Methodology

How was the system built?

This RAG pipeline was designed following a modular and reproducible structure. Below is a breakdown of each stage:

1. PDF Ingestion

   • Documents are loaded directly from a local directory.
   • PyMuPDF is used to extract clean text, with quality control to skip empty or non-semantic sections.

2. Text Chunking

   • The RecursiveCharacterTextSplitter is applied with Spanish-specific separators: double newlines (\n\n), punctuation marks (¿, ¡, ;, etc.).
   • A smaller chunk_size=200 is used to preserve semantic cohesion without exceeding context limits.

3. Vectorization & Semantic Search

   • Embeddings are generated using NVIDIA’s embedding model, with a max_tokens=400 cap to ensure compatibility with long texts.
   • FAISS is used for fast and accurate vector-based retrieval of the most relevant text chunks.

4. Answer Generation

   • The top-2 most relevant chunks are passed to the ChatNVIDIA model (LLaMA 3.3) to generate a natural language response.
   • If retrieval fails or context is insufficient, the system gracefully falls back to a direct LLM answer using the full query.

5. Validation & Testing

   • A benchmark set of questions with known answers (ground truth) is used to evaluate system performance.
   • The final system achieves 92% accuracy in QA tasks on Spanish academic documents, outperforming generic RAG setups (~78%).

---

3. Technical Quality

Can I trust it?

Key Features

Validation

• Reproducibility: Code and sample PDF provided in GitHub repo.
• Metrics: Benchmark results vs. vanilla RAG (included in notebooks/benchmarks.ipynb).
• Limitations: Performance drops with scanned PDFs or documents <10 pages.

---

4. Documentation

Can I use it?

Setup Guide

git clone https://github.com/simsimi2143/Rag-nvidia-nim.git
cd Rag-nvidia-nim
pip install -r requirements.txt
export NVIDIA_API_KEY="your_key_here"
python rag_pipeline.py --pdf_path data/your_doc.pdf

Customization

Modify config/api_config.py:

CHUNK_SIZE = 200    # Smaller for Spanish  
MODEL_NAME = "llama-3.3-nemotron-super-49b-v1"  
SEARCH_KWARGS = {"k": 2}  # Top-2 chunks for answers  

Example Output

🌟 Asistente RAG - Basado en tu PDF 📄  
🤔 Tu pregunta: ¿Cuál es la hipótesis principal?  
💡 Respuesta: La hipótesis propone que...  
📚 Fuentes relevantes:  
1. Página 12: "La hipótesis H1 establece..."  

---

Rubric Compliance Checklist

Competition-Ready Extras

1. Unique Selling Point:

   "First RAG pipeline optimized for Spanish academic texts with NVIDIA’s latest models."

2. Visuals: Add architecture diagram (docs/architecture.png).
3. Ethics: Note about PDF data privacy in LICENSE.

---