Smart Search with ChatGPT: 8-Wk Proof of Concept

DESCRIPTION: The system will encompass a knowledge base (KB) containing approximately one hundred documents pertinent to a specific business domain. The Knowledge Base in this solution will primarily consist of PDF documents manually uploaded to Azure Storage with no initial integrations with source systems. It is important to note that this architecture focuses on machine-readable documents, eliminating the need for Optical Character Recognition (OCR) services. Additionally, all documents will be in the English language, excluding the optional translation component shown in the architecture. Azure Form Recognizer will be utilized to extract textual components, such as paragraphs, from the given documents, resulting in structured and readily usable information. This textual information will then be transposed into vector format using embeddings, with the assistance of the dedicated OpenAI service. The vectorized information will be stored in a non-relational database, such as Redis or CosmosDB, creating an effective repository of paragraphs ready for retrieval. A similar process will be applied to user-provided questions in natural language. By employing embeddings, these questions will be transformed into vector format to facilitate search purposes. The vectorized question will be compared with the content of the paragraphs extracted from the documents to identify the most relevant paragraphs and documents to construct a response. Using a "Top-K" logic, the resulting paragraphs will be filtered, and the ChatGPT service will be engaged to create a summary of their content, which directly answers the original user's question. Azure OpenAI services, including ChatGPT, will be harnessed for content retrieval and the generation of responses in natural language. Language support is crucial, with an initial focus on English for both source documentation and user-submitted questions. During the implementation phase, the potential for supporting other languages, at least for user queries, will be assessed. Leveraging Azure features, the system will be capable of responding to specific queries, including requests for summaries of research conducted by a specific institute on a particular topic and precise inquiries related to specific subjects. A user-friendly interface is essential, allowing users to input questions and receive model-generated responses in textual form. This interface will be made accessible through a prototypical web application. Initially, the system will be deployed on the Azure Avvale subscription, with plans for migration to the Azure client subscription when available.

OUTCOMES:

• Contextual understanding of the selected Azure analytical tools
• Discover how ML and AI could be exploited to improve the considered process
• Model explanation to extract meaningful information from the data

DELIVERABLES:

• Setup Azure prototypical reference architecture
• Knowledge base from manually uploaded docs (pdf)
• Azure ML and AI overview (Form recognizer, OpenAI services)
• Simplified graphical interface for user interaction
• Next steps plan to extend the POC capabilities to other business scenarios

PLANNING:

• Week 1-3: Briefing about the POC scope and high-level details of setting up the Azure analytical ecosystem. Data selection and understanding session with the business team. Architectural setup.
• Week 3-8: Set up of AI and ML environment. Onboard datasets and solution development. Model test and fine-tuning.
• Week 8: Roadmap based on POC results and insights

AUDIENCE: ICT Dept., Data Scientists, Business users

LANGUAGES: Italian or English