An AI-Powered Support Companion for Drug Awareness in Saudi Arabia.

Baseera is a Retrieval-Augmented Generation (RAG) system built to provide empathetic, evidence-based guidance grounded in official Saudi Ministry of Health (MOH) and NCNC protocols. By utilizing open-source LLMs, this project demonstrates strong engineering practices in local data privacy, open-source AI orchestration, and grounded response generation.

The following diagram illustrates the upgraded RAG lifecycle using Semantic Chunking:

graph TD
    A[PDF Data Sources] -->|PyPDFDirectoryLoader| B[ingest.py]
    B -->|SemanticChunker| C{Vector Store}
    C -->|HuggingFaceEmbeddings| D[FAISS Index]

    E[User Query] -->|Streamlit UI| F[app.py]
    F -->|Similarity Search| D
    D -->|Relevant Context| G[Llama 3.1 8B Instruct]
    G -->|Grounded Response| H[User]

• LLM: Llama 3.1 8B Instruct (via Hugging Face Inference API)
• Orchestration: LangChain (Community & Partner Packages)
• Vector Database: FAISS (Local)
• Embeddings: sentence-transformers/all-MiniLM-L6-v2
• UI: Streamlit
• Data Sources: Official MOH & NCNC PDFs

Create a .env file in the project root and add your Hugging Face token:

HUGGINGFACEHUB_API_TOKEN=your_token_here

Install the experimental package required for semantic processing:

pip install langchain-experimental

# Clone the repository
git clone https://github.com/your-username/baseera-ai.git
cd baseera-ai

# Create and activate virtual environment
python -m venv venv
source venv/Scripts/activate  # Windows
# source venv/bin/activate    # macOS/Linux

# Install dependencies
pip install -r requirements.txt

Before running the application, process the PDF documents inside the data/ folder to generate the local FAISS vector store:

python ingest.py

What makes this version (v1.0.0) different?

• Intelligence: It no longer splits text every 600 characters. Instead, it uses HuggingFaceEmbeddings to determine when a topic or meaning changes, keeping medical protocols intact.
• Precision: By ensuring sentences aren't cut in half, the Llama 3.1 model has cleaner context to work with, significantly reducing the "ugly" character hallucinations seen in earlier versions.

streamlit run app.py

• Semantic Document Processing: Utilizes AI-driven chunking to preserve the integrity of Arabic medical guidelines, ensuring retrieval occurs at the "thought" level rather than the "character" level.

• Grounded Citations: Every response begins with "Based on [Source]…" to ensure transparency and trust.

• Localized RTL Support: Dynamic UI rendering that automatically detects Arabic text and aligns it Right-to-Left (RTL) for a native user experience.

• Automatic Directory Ingestion: Any new PDF added to the data/ folder is picked up during the next ingestion run.

• System Prompt Engineering: Specialized instructions ensure:

  • Empathetic, non-judgmental tone
  • Medical disclaimers
  • Cultural and regional sensitivity

• Local Vector Store: No external database required — all embeddings are stored locally.

sequenceDiagram
    participant U as User
    participant S as Streamlit App
    participant V as FAISS Index
    participant L as Llama 3.1

    U->>S: Asks about drug prevention
    S->>V: Search for relevant document chunks
    V-->>S: Return top-k matching segments
    S->>L: Context + System Prompt + Query
    L-->>S: Grounded response with citations
    S-->>U: Final Answer

While Baseera is a functional RAG MVP, the following enhancements would elevate it to a production-grade system, particularly for the Saudi market.

• Semantic Chunking for Arabic
  Transition from character-based splitting to semantic chunking using embeddings trained on Arabic (e.g., AraBERT, MARBERT) to prevent medical or legal concepts from being split mid-sentence.

• Morphological Analysis
  Integrate Arabic NLP tools such as CAMeL Tools for lemmatization and diacritic normalization, improving retrieval accuracy across different word forms.

• Dialect-Aware Prompting
  Enhance the system prompt to better interpret Saudi dialects (Najdi, Hejazi, Southern, etc.) while maintaining responses in formal Fusha.

• Substance Slang Recognition
  Improve understanding of local slang related to substances without reproducing it in responses, preserving professionalism and safety.

• Hybrid Retrieval (BM25 + Vector Search)
  Combine keyword-based search with vector similarity to ensure high-precision retrieval for medical terms, legal references, and protocol identifiers.

• Reranking Layer
  Introduce a cross-encoder reranker (e.g., BGE-Reranker) to re-score the top-k FAISS results and ensure the most medically and culturally relevant context is selected.

• Automated RAG Metrics
  Integrate the RAGAS framework to continuously evaluate system quality using metrics such as:

  • Faithfulness: Is the response grounded in the source PDFs?
  • Answer Relevance: Does it directly address the user’s concern?
  • Context Precision: Is the retrieved context truly useful?

• Regression Testing for Prompts
  Track performance changes when updating prompts, models, or retrieval strategies.

Baseera is an educational and awareness tool. It does not replace professional medical, psychological, or legal advice. For emergencies or medical decisions, always consult licensed professionals or official Saudi health authorities.

This project is released for educational and research purposes.