🚀 Project Overview
This project is a fully functional Retrieval-Augmented Generation (RAG) Assistant built as part of the AAIDC Module 1 Certification. While the initial template provided the skeleton, I implemented the "muscles" of the system—logic for document loading, chunking, embedding, and vector retrieval.

Going beyond the basic requirements, I enhanced the application by replacing the command-line interface with a Streamlit Web UI. This transforms the project from a script into a user-friendly product where users can visually manage their knowledge base and chat with their PDF documents in real-time.

🌟 Key Features
interactive Web Interface: A clean, chat-like UI built with Streamlit that replaces the standard terminal loop.

Universal PDF Loader: Automatically scans, loads, and parses PDF files from the data/ directory using pypdf.

Smart Context Retrieval: Uses ChromaDB and Sentence Transformers to semantically search for the most relevant document chunks.

Multi-Model Intelligence: The system is agnostic to the LLM provider. It seamlessly switches between:

OpenAI (GPT-4o)

Groq (Llama 3.1)

Google (Gemini 2.0 Flash)

Robust Error Handling: Specifically engineered to handle version conflicts, ensuring stability on Python 3.11/3.12 environments.

🛠️ Technical Implementation

1. Document Ingestion & Chunking
   The system reads raw PDF files and processes them using the RecursiveCharacterTextSplitter. This is superior to simple whitespace splitting because it respects sentence boundaries and paragraph structure, ensuring that the semantic meaning of the text is preserved when fed into the AI.

Code Snippet: Chunking Logic

text_splitter = RecursiveCharacterTextSplitter(
chunk_size=500,
chunk_overlap=50,
length_function=len,
separators=["\n\n", "\n", " ", ""]
)
2. Vector Database & Embeddings
I utilized ChromaDB as the vector store. When documents are loaded, they are converted into dense vector embeddings using the sentence-transformers/all-MiniLM-L6-v2 model. This allows the system to perform "Semantic Search"—finding text that matches the meaning of the user's question, not just the keywords.

3. The RAG Pipeline
   The query pipeline follows a strict flow to minimize hallucinations:

Retrieve: The user's question is embedded, and the top 5 most similar chunks are fetched from the database.

Augment: These chunks are combined into a single "Context" block.

Generate: A strict system prompt instructs the LLM to answer only based on the provided context.

🚧 Challenges & Solutions
One of the significant hurdles I faced during development was the "Dependency Civil War" between modern AI libraries.

The Issue: The latest Python 3.14 release is not yet supported by onnxruntime (required by ChromaDB), while newer LangChain versions required Pydantic v2, clashing with older database dependencies.

The Solution: I engineered a specific requirements.txt and enforced a Python 3.11 environment. This resolved the symbol compatibility issues and allowed pydantic-settings to function correctly with the latest LangChain orchestration.

💻 How to Run This Project
Prerequisites: Python 3.11 and an API Key (OpenAI, Groq, or Google).

Clone the Repo:
git clone https://github.com/HEPTA-111/rt-agentic-ai-certification.git

Install Dependencies:
pip install -r requirements.txt

Run the App:
streamlit run src/app.py

🔮 Future Improvements
Upload Button: Adding a drag-and-drop file uploader directly in the UI (currently it loads from a folder).

Chat History Persistence: Saving chat logs to a local file so conversations aren't lost on refresh.

Source Citation: Modifying the prompt to strictly cite which document file provided the answer.