NEXA AI Experience: RAG-Powered Knowledge Base 🚗

A high-performance, terminal-based Retrieval-Augmented Generation (RAG) system for Maruti Suzuki NEXA vehicle manuals. This project implements a Hybrid Search Architecture to deliver precise technical answers from complex PDF documentation, specifically optimized for tabular data (e.g., tyre pressures, maintenance schedules).

🏗 System Architecture

This project solves the "Retrieval Fragmentation" problem common in RAG systems by using a two-stage filtering process.

graph LR
    A[User Query] --> B[Hybrid Retrieval]
    B --> C{Pinecone Vector DB}
    C -->|Fetch Top 20 Chunks| D[Cross-Encoder Reranker]
    D -->|Filter Top 5 Contexts| E[LLM Inference]
    E --> F[Answer Generation]

Loading

The Pipeline

1. Ingestion Engine (ingest.py):
   • Uses Docling with RapidOCR for layout-aware parsing.
   • Converts PDFs to semantic Markdown, preserving table structures (critical for technical specs).
2. Indexing (index.py):
   • Adaptive Chunking: Implements a hybrid strategy (Semantic Splitting + Recursive Fallback) to handle large context windows without exceeding Pinecone's 40KB metadata limit.
   • Embeddings: Uses BAAI/bge-small-en-v1.5 (SOTA open-source model).
3. Inference Engine (chat.py):
   • Retrieval: Fetches a wide net (20 chunks) to ensure tables aren't missed.
   • Reranking: Uses ms-marco-MiniLM-L-6-v2 to grade relevance and filter noise.
   • Generation: Powered by OpenAI GPT-OSS-20B (via Groq) for sub-second latency and strong reasoning capabilities.

🚀 Key Features

• Multi-Model Support: Unified knowledge base for Jimny, Grand Vitara, Fronx, and Baleno.
• Table-Aware Inference: Accurately extracts data from complex Markdown tables (e.g., "Full Load" vs. "Normal Load" tyre pressures).
• Safety Protocol: System prompt engineered to prioritize "WARNING" and "CAUTION" blocks.
• Zero-Latency CLI: Optimized for terminal usage with streaming responses.

📦 Installation & Setup

1. Clone the Repository

   git clone https://github.com/yourusername/NexaMs-RAG-System.git
   cd NexaMs-RAG-System

2. Install Dependencies

   pip install -r requirements.txt

3. Configure Environment Create a .env file in the root directory:

   PINECONE_API_KEY=your_pinecone_key
   GROQ_API_KEY=your_groq_key

4. Ingest Data (Optional) Place PDFs in data/raw_pdfs/ and run:

   python ingest.py
   python index.py

5. Run the Engine

   python chat.py

👨‍💻 Tech Stack

• Orchestration: LlamaIndex
• LLM: OpenAI GPT-OSS-20B (Groq)
• Embeddings: BAAI/bge-small-en-v1.5
• Reranker: sentence-transformers (cross-encoder)
• Vector Database: Pinecone (Serverless)
• OCR: Docling + RapidOCR

---