ALLAM-based-Retrieval-Augmented-Generation-Arabic-Conversational-Alzheimer-Assistance

ALLaM-RAG: A Context-Aware Saudi Arabic Conversational AI System for Alzheimer’s Support

Overview

ALLaM-RAG is a Saudi Arabic conversational AI system that combines Retrieval-Augmented Generation (RAG) with the ALLaM Large Language Model to provide context-aware assistance for Alzheimer's patients and caregivers.

The system retrieves relevant information from specialized knowledge bases using hybrid vector search and reranking before generating accurate, personalized responses through the ALLaM language model.

The system supports two knowledge domains: Alzheimer's patient support and general Saudi Arabic knowledge.

---

Features

• Saudi Arabic conversational assistant
• Retrieval-Augmented Generation (RAG) architecture
• Powered by the ALLAM Large Language Model
• Hybrid semantic and keyword retrieval using Weaviate
• Multilingual reranking using Cohere Rerank
• Context-aware response generation
• Alzheimer's patient support
• General Saudi knowledge support
• Real-time time and day awareness
• Arabic text normalization and intent detection
• Query classification and knowledge-base selection
• Retrieval and generation evaluation
• Comparison between RAG and non-RAG configurations

---

System Architecture

User Query
      │
      ▼
Arabic Text Normalization
      │
      ▼
Query Classification
      │
      ▼
Knowledge Base Selection
      │
      ▼
Weaviate Hybrid Search
      │
      ▼
Cohere Reranker
      │
      ▼
Relevant Context
      │
      ▼
ALLaM Language Model
      │
      ▼
Generated Response

---

Technologies Used

• Python
• ALLaM-7B-Instruct
• Transformers
• Hugging Face
• Weaviate Vector Database
• Cohere Rerank API
• NumPy

---

Knowledge Bases

The system utilizes two specialized knowledge bases.

Alzheimer's Knowledge Base

The Alzheimer's knowledge base is built from a Saudi Arabic conversational dataset focused on patient-oriented interactions and support.

It contains information related to:

• Patient identity
• Memory assistance
• Medication reminders
• Daily activities
• Alzheimer's symptoms
• Caregiving guidance
• Orientation support
• Health-related conversational scenarios

The original source dataset is the Saudi-Arabic-Alzheimers-Conversational-Dataset-Parameterized, containing parameterized Saudi Arabic question-answer examples. The dataset was a part of the graduation project: AlzCare in Taibah University.

General Knowledge Base

The GeneralQA knowledge base provides Saudi Arabic information across general domains, including:

• Islamic knowledge
• Saudi Arabia
• Culture
• Geography
• Public information

Separating the knowledge bases allows the system to retrieve information from the domain most relevant to the user's query.

---

Retrieval Pipeline

The retrieval pipeline consists of:

1. Query preprocessing and Arabic text normalization
2. Query classification
3. Knowledge base selection
4. Hybrid retrieval from Weaviate
5. Cohere multilingual reranking
6. Relevant context selection
7. Response generation using ALLaM

The retrieval stage combines semantic and keyword-based search before reranking the retrieved candidates to identify the most relevant context for generation.

---

Evaluation

The system was evaluated using question-answer examples from both the AlzheimerQA and GeneralQA datasets.

The evaluation measures both retrieval quality and the quality of the final generated answers.

Retrieval Evaluation

These metrics evaluate how effectively the retrieval pipeline identifies relevant information and ranks it among the retrieved results.

RAG Answer Evaluation

The generated responses were evaluated using RAGAS-based metrics:

The results indicate that the retrieved context generally provides relevant and sufficiently grounded information for response generation.

Model Comparison

The RAG-based system was also compared with alternative configurations:

The comparison demonstrates the contribution of retrieval to the overall system performance, with the ALLaM-RAG configuration achieving the highest evaluation score among the tested configurations.

---

Query Classification Evaluation

The query classification component was evaluated separately to determine whether incoming questions were correctly classified before selecting the corresponding knowledge base.

The classifier achieved:

• Accuracy: 92%
• F1 Score: 0.93

This classification step supports routing between the Alzheimer's and GeneralQA knowledge bases.

---

Evaluation Datasets

The evaluation data contains examples from both system domains:

Evaluation Dataset
│
├── AlzheimerQA
│   └── alz_* question IDs
│
└── GeneralQA
    └── GeneralQA examples

The evaluation covers both retrieval and answer generation rather than evaluating only the final generated response.

---

Example Queries

Alzheimer's Support

• مين أنا؟
• وين ساكن؟
• كيف حالتي الصحية؟
• نسيت أخذ الدواء
• ليه أنا هنا؟

General Saudi Arabic Knowledge

The system can also respond to general Saudi Arabic questions using the GeneralQA knowledge base.

---

Project Structure

app.py
chatbot.py
rag_pipeline.py
weaviate_server.py
requirements.txt
README.md

---

Running Locally

Install dependencies:

pip install -r requirements.txt

Configure the required environment variables:

• COHERE_API_KEY
• HF_TOKEN
• WEAVIATE_URL (if applicable)
• WEAVIATE_API_KEY (if applicable)

Launch the application:

python app.py

---

System Architecture

---

Deployment

The project is designed for deployment using:

• Hugging Face Models
• ALLaM language model
• Weaviate Vector Database
• Cohere Rerank API

The model and project resources are hosted through Hugging Face, while Weaviate provides the retrieval infrastructure and Cohere provides multilingual reranking.

---

Demo

Open the ALLAM-RAG website

---

Key Results

The evaluation demonstrates the effectiveness of combining retrieval with ALLaM generation:

• 82.7% overall RAG evaluation score
• 91.1% faithfulness
• 88.2% context recall
• 86.8% Recall@5
• 81.2% Precision@5
• 92% query classification accuracy
• 0.93 classification F1 score

The ALLaM-RAG configuration achieved 82.7%, compared with 61.4% for ALLaM without RAG.

---

Future Improvements

• Multi-turn conversation memory
• Expanded Alzheimer's knowledge base
• Further retrieval and generation optimization

---

Author

Shahad Aljohani

Computer Science Graduate

---

Privacy Notice

This repository contains a public research version.

Private implementation details, API keys, deployment configurations, and some internal pipeline components are excluded.

The complete implementation is maintained privately.