Project Overview

The AI Learning Assistant (peped-BE) is the AI-powered backend backbone developed for Amartha’s Learning Management System (Amarthapedia). The core challenge was managing vast, scattered internal knowledge bases and transforming them into a responsive, intelligent assistant. This project focuses on building an efficient Retrieval-Augmented Generation (RAG) system to ensure every employee (the A-Team) can access accurate, highly relevant information instantly through natural conversation.

1. Strategic Objectives

The primary goal of developing this intelligent assistant was to democratize knowledge access within the company. Key strategic objectives included:

  • Enhance Knowledge Accuracy: Ensure the assistant strictly grounds its answers in valid internal data to effectively minimize AI hallucinations.
  • Optimize Response Speed: Reduce technical friction and retrieval latency so the learning flow is never interrupted by long wait times.
  • Empower On-Demand Learning: Provide the A-Team with instant, 24/7 learning support, helping them overcome study blockers and understand complex materials independently without waiting for human intervention

2. Engineering Strategy & Architecture

To build a highly reliable and intelligent assistant, I utilized an Agentic RAG approach backed by a modern, high-performance tech stack:

  • Agentic Orchestration (LangGraph & FastAPI): Leveraged FastAPI for high-performance APIs and LangGraph to manage complex AI workflows, enabling the assistant to "think" and reason before generating an answer.
  • Advanced Retrieval (Qdrant & LightRAG): Implemented Qdrant as the primary vector database for precise semantic search, alongside research into LightRAG to improve the accuracy of information retrieval from dense knowledge bases.
  • Performance Engineering (Redis Caching): Utilized Redis for semantic caching, which significantly reduces the LLM workload and accelerates response times for frequently asked or similar questions.
  • Self-Hosted Infrastructure: Prioritized data security and cost-efficiency by deploying services like Langfuse and rerankers on a local Proxmox server using Docker.

Summary of Core Features

  • Semantic Search & Retrieval: The assistant's ability to understand the deeper context and intent of a user's query, moving far beyond simple keyword matching.
  • Long-Term Agent Memory: Integration of agent-managed long-term memory (via Qdrant) to provide a continuous, highly personalized learning experience.
  • Seamless LMS Integration: A robust backend architecture designed to integrate smoothly with the Amarthapedia frontend, ensuring flawless data transition between systems.