IslamWeb RAG Chatbot

Project summary

Provide useful Arabic answers from an established religious knowledge base without allowing an automated assistant to overstep when a question requires contextual judgment from a scholar.

System view Sanitized

01 User query

->

02 Intent agent

->

03 RAG response

->

04 Risk check

->

05 Scholar handoff

Context

Developed for the Qatar Ministry of Awqaf and Islamic Affairs as part of its IslamWeb services, the assistant needs a clear boundary between questions that can be answered from retrieved material and questions that require a scholar’s interpretation. The product therefore needed both an effective RAG path and a first-class escalation path.

The goal was not to maximize automated resolution at any cost. It was to create a system that could be helpful while recognizing when automation was no longer appropriate.

Constraints

• Arabic queries may be colloquial, incomplete, or context dependent.
• Retrieved evidence must come from the approved fatwa corpus.
• Sensitive or ambiguous requests must reach a qualified human.
• Conversation state must survive the transition from assistant to scholar.
• The experience needed to operate across web and Telegram channels.

My role

I developed the Arabic RAG assistant, generated embeddings with Gemini models, implemented BigQuery vector retrieval, and created the intent-detection agent responsible for identifying conversations that need escalation.

I also integrated the assistant with Telegram as an additional delivery channel.

Retrieval and routing design

The normal path retrieves semantically relevant rulings and uses them as context for an answer. A separate intent layer examines the request and conversation state for signals that the assistant should defer.

When escalation is required, the system routes the conversation to a Mufti rather than producing an unjustifiably confident response. Treating human handoff as a product outcome, rather than an exception, shaped the architecture.

Evaluation and safeguards

Evaluation needs to cover two distinct behaviors: answer quality when retrieval is appropriate and routing quality when human judgment is required. A system can retrieve well and still be unsafe if it fails to defer.

Public details intentionally stop short of exposing internal routing categories or thresholds. The architectural principle is straightforward: uncertainty and sensitivity must be represented in the system’s control flow.

Production deployment

The assistant uses Gemini-generated embeddings and BigQuery vector search within a Google Cloud environment. Channel integration is separated from retrieval and routing logic so the same core behavior can support multiple interfaces.

Results and lessons

The system delivered an Arabic domain assistant with an explicit scholar-escalation mechanism and Telegram integration.

The key lesson was that human-in-the-loop design should begin with the handoff contract: what triggers escalation, what context travels with the request, and how the automated system communicates its limits.