🏡 Rag_Library_AI

This project is a Retrieval-Augmented Generation (RAG) application designed to act as an intelligent librarian. By indexing technical Python books, it allows users to ask complex questions and receive answers grounded in the specific text of those books, complete with context.

---

📸 Screenshots

UI of the Rag Library Dashboard

---

🚀 Key Features

• Dual Interfaces: Interaction via a clean Streamlit Web Dashboard (app.py) or a lightweight CLI Tool (query.py).

• Rust-Powered Performance: Document ingestion uses a custom high-speed text splitting package (fast_chunker) built with Rust and exposed to Python using PyO3.

• Semantic Guardrails: Every response undergoes a real-time Groundedness Audit using local embeddings (all-MiniLM-L6-v2) and Cosine Similarity to evaluate if the AI's response is mathematically justified by the source texts.

• Deterministic Safety Fallbacks: Completely blocks hallucinated library syntax or jumbled text blocks, gracefully bypassing checks only when the AI correctly identifies a lack of context.

---

📁 Project Structure

The project architecture is structured for scalability, featuring isolated scanning, processing, and documentation assembly engines.

rag-technical-app/
├── fast_chunker/          # Custom Rust text splitting module
│   ├── src/
│   │   └── lib.rs         # PyO3 Rust chunking implementation
│   ├── Cargo.toml         # Rust dependency definitions
│   └── pyproject.toml     # Rust build & metadata configuration
├── vectorstore/
│   └── db/                # Persistent ChromaDB vector storage files
├── screenshots/           # Screenshots of the Rag App
├── app.py                 # Core Streamlit Web Application interface
├── query.py               # Command-line (CLI) interaction utility
├── ingest.py              # PDF Parsing and Vector Store Ingestion engine
├── guardrail.py           # Semantic Guardrail verification system
├── requirements.txt       # Python environment dependencies
├── tests/                 # Pytest suite
└── .gitignore             # Git exclusion parameters  

---

🚀 Getting Started

Prerequisites Python 3.11 or higher

A Google Gemini API Key (obtained via Google AI Studio)

Installation & Environment Setup

Clone the repository and move into your project root:

git clone <https://github.com/reory/rag_library_ai>
cd rag-technical-app

Create and activate a virtual environment:

# Windows
python -m venv venv
venv\Scripts\activate

# macOS / Linux
python3 -m venv venv
source venv/bin/activate

Install the project dependencies.

This automatically installs the specialized, pre-compiled fast_chunker Linux/Windows target wheel along with core frameworks:

uv pip install -r requirements.txt

Environment Variables Configuration

• Create a .env file in the root directory to store your API credentials locally:

Code snippet

• GOOGLE_API_KEY=your_actual_gemini_api_key_here

Ingesting Documents

Run the ingestion pipeline to compile the high-speed Rust chunks into your persistent vector store:

python ingest.py

---

💻 Usage

• Web Interface (Streamlit Dashboard) Launch the responsive, styled dashboard interface locally:

python -m streamlit run app.py

• Open http://localhost:8501 in your browser.

The UI features styled badge elements for core Python libraries, interactive querying, a quick history clear utility, and full audit logs mapping your Guardrail Verification parameters.

---

🛡️ How the Hallucination Guardrail Works

• Every response undergoes a real-time verification pipeline before reaching the interface:

• Extraction & Generation: The system retrieves the top 3 document chunks and prompts Gemini 2.5 Flash to answer using only that context.

• Semantic Auditing: A local SentenceTransformer maps both the raw context and the AI's response into dense vector arrays.

• Similarity Check: A Cosine Similarity calculation evaluates factual alignment. If the score drops below 0.62, the response is immediately blocked.

• Safe Refusal Pass: Standard phrases admitting a lack of context (e.g., "The answer isn't in the context") bypass the math filter with a perfect 1.0 score.

⚙️ Parameter Tuning

Adjust the strictness of the filter inside guardrail.py:

class RAGGuardrail:
    def __init__(self, threshold=0.60):

• Drop to 0.55 if blocking valid text; raise to 0.65+ if hallucinations slip through 😊

---

🛣️ Roadmap Features

• High-speed Rust text chunking integration (fast_chunker).
• Core Streamlit Web UI and CLI query processing pipelines.
• Mathematical Semantic Guardrail Filter (Cosine Similarity Verification).
• Self-Correction Loop: Automatically re-prompt Gemini to fix answers if the guardrail filter detects a hallucination.
• Conversational Memory: Store message history in Streamlit session state to support continuous multi-turn chat dialogues.
• Source Citation: Extract and display document metadata (Book Title, Page Numbers) immediately below verified answers.
• Semantic Input Routing: Block or redirect out-of-scope user prompts upstream before triggering database or LLM resources.

---

• Built by Roy Peters 🙂