The primary objective of the knowledge management digital assistant is to create an intelligent, efficient, and user-friendly solution that enables users to retrieve precise information from diverse sources of content.

Customer is in the business of providing energy & sustainability solutions globally. They currently have a web portal for knowledge management, where they have lot of content (machine readable and scanned documents, html pages, images, videos, audios etc.) related to green ports and carbon emissions that enable policy making and regulations to achieve sustainability goals. Though the portal has search facility, it takes lot of time to get required information easily, at times lot of unwanted information comes out of search, some of the content is scanned and images that makes it difficult to search.

So, customer wanted a question & answering digital assistant that can provide precise answers quickly for natural language questions using the content across knowledge management portal.

Designed and deployed a GenAI solution that meets the objective and addresses challenges being faced by the customer currently. The digital assistant that sits on existing knowledge management portal, provides accurate answers quickly for various questions.

Digital assistant takes questions both in text and voice, and answers also come in both text and voice.

It can answer questions whose answers lie in scanned documents and images as well. It can answer statistical and forecasting questions also, provided sufficient data is there in the portal.

For every answer, the assistant also provides references to original documents/web pages from where it picked up the answer.

It also allows users to upload a document and ask questions to be answered, in real time, using the content in the document.

The system employs a hybrid architecture that uses AI-driven search and real-time query processing. Document and image processing starts with Optical Character Recognition (OCR), while audio is converted to text using Whisper. The API is built with FastAPI, and Retrieval Augmented Generation (RAG) is used to retrieve relevant chunks from the vector database (Qdrant). These chunks are then passed to the LLM GPT-4o-mini to generate summarized and formatted responses1. Langchain is used for memory storage. The response is streamed via WebSocket. Secure communication is ensured through Uvicorn with SSL, while temporary file uploads are managed within a separate Qdrant collection, allowing on-demand content retrieval with automatic data deletion.

The frontend is developed using HTML, CSS, and JavaScript, providing a user-friendly chat interface. It supports structured responses with source attribution, file uploads for document-based search, and an interactive experience for engaging with AI-powered responses.

The backend application is built using Python and FastAPI, managing API interactions, data processing, and integration with Qdrant for vector-based search. It processes user queries in real-time, retrieves relevant content using Retrieval Augmented Generation (RAG), and generates AI-driven responses via GPT-4o-mini. The system also incorporates WebSocket communication for seamless response streaming and efficiently manages temporary data for document-based queries, ensuring smooth user interactions.

The implementation of the knowledge Management assistant has significantly reduced manual search time by up to 70%. Previously, users had to spend valuable time manually searching through articles, research papers, and news within the portal—often dealing with scanned PDFs and images, making the process even more challenging.