Makhfi AI is a source-supported assistant over Furqan Qureshi Blogs, a research platform covering the Quran, Hadith, modern science, archaeology, and ancient history. It helps users explore the platform's video and text library through natural-language questions, source context, and supported video summaries.
The interface can respond in Urdu while keeping the conversation connected to material from the platform.
The Knowledge Problem
The platform contained useful material across individual videos and written sources. Finding the relevant part required knowing where to look and manually moving through the library. The product needed a conversational path into that material without presenting generated text as an independent authority.
What I Built
- A Next.js chat experience with conversation history and bilingual interaction.
- A FastAPI runtime for messages, sessions, retrieval, and source-aware response generation.
- A LangGraph workflow coordinating retrieval, model reasoning, and supported tools.
- Pinecone-backed semantic retrieval across the prepared knowledge library.
- Source context that connects an answer to relevant platform material.
- Supported video summarization flows.
- Persistent conversation and metadata storage with SQLModel and PostgreSQL through Supabase.
- Long-term memory through LangMem, exposed as a user-visible part of the experience.
- Supabase authentication and JWT-based session handling.
History, retrieved context, and generated responses are presented as parts of one continuing research session.
Decisions That Mattered
Keep source context visible
The subjects covered by the platform deserve traceability. A generated answer is a navigation and synthesis layer, so the interface leads users back toward the material supporting the response.
Make retained memory a product surface
Long-term memory can make later conversations more coherent, but retained context should not be invisible. The system treats memory as explicit application state rather than an unlimited transcript passed into every request.
Separate platform knowledge from general model language
The model can write and reason broadly; the retrieval layer identifies what the platform itself contains. Keeping that distinction clear supports better source behavior and more useful no-evidence handling.
The Delivered Experience
Makhfi was deployed for a small internal user group with chat history, Urdu and English interaction, source-supported responses, persistent memory, and supported video-summary workflows.
What I Took Forward
Knowledge assistants become useful when people can inspect where an answer came from and control the context carried forward. Retrieval quality, citation behavior, and memory boundaries are product decisions—not backend details.
For the broader design and evaluation approach, see agentic RAG for company knowledge and multimodal document AI in 2026.
