Gen AI-powered document generation

Our client, a major industrial automation company, provides industrial automation solutions for customers in the energy sectors, including oil and gas, power and water, and chemicals, through long-term projects at customer sites.

They faced significant challenges in generating Functional Design Specification (FDS) documents. These documents, essential for project implementation, were time-consuming and complex, often exceeding 200 pages with numerous subsections of unstructured data. The manual process required extensive effort, taking over 80 hours per document, leading to inefficiencies, long turnaround times, and potential project delays. Additionally, manual intervention increased the risk of errors and inconsistencies, resulting in operational delays, higher costs, and resource strain on the engineering team.

During the discovery phase with SMEs, the client identified several key challenges in automating the FDS document generation process. The high volume and complexity of FDS documents, which often span hundreds of pages and include diverse content types like tables, images, and text, made manual extraction and processing slow and error prone.

Additionally, data inconsistency was a significant barrier, as FDS data came from various unstructured sources such as historic FDSs, BoMs, technical proposals, and product manuals. This lack of standardization led to errors and inconsistencies, making it difficult to maintain data accuracy across numerous subsections.

Furthermore, integrating different data types into a coherent document required advanced AI and document processing techniques. These challenges resulted in inefficiencies, prolonged timelines, and inconsistent quality, underscoring the need for an automated solution to streamline the process.

To address these challenges, Fractal developed a Gen AI-powered document generation solution using Azure AI services and Large Language Models (LLMs) like GPT-4. The solution automates data extraction, processing, and generation of FDS documents, improving speed, consistency, and scalability.

It includes three primary layers: Data extraction, FDS document building, and user interface (UI).

1. Data extraction layer: This layer uses Azure Document Intelligence to extract structured and unstructured data from various sources, including Excel files, IP schedules, BoMs, legacy FDS PDFs, and technical proposal documents stored in Azure Blob Storage and SQL databases.

A custom AI extraction layer applies Optical Character Recognition (OCR) and advanced document processing models to accurately capture tables, images, and text, ensuring the extraction of all relevant data for further processing. Extracted data is indexed in Azure AI Search, facilitating efficient and semantic data retrieval for subsequent document generation tasks.

2. FDS document building layer: This layer employs semantic matching to retrieve content from the indexed data, ensuring contextually relevant and accurate inclusion in the FDS document. Azure OpenAI (GPT-4) is used to refine the content, generating contextually accurate and consistent text for each section of the FDS.

This multi-stage processing approach includes content refinement, result validation, and iteration, ensuring that the content is accurate, coherent, and in line with the document’s specifications before generating the final output. The finalized FDS document is automatically generated in Word format, with all tables, images, and text properly formatted and inserted into the appropriate sections.

3. UI layer: This layer provides a user-friendly interface for users to easily upload input files and trigger the FDS generation process. The UI integrates seamlessly with Azure Blob Storage for managing input files and Azure Functions to initiate and orchestrate the backend processing tasks. Users receive real-time feedback on the document generation process and can easily access the generated FDS documents.

Early observations indicate that the automated RAG framework has significantly improved the consistency and accuracy of the FDS documents, minimizing variations and ensuring high-quality outputs. The solution has drastically reduced the document generation time from over 60 hours to just 30 minutes, accelerating project timelines and boosting overall efficiency. This improvement has enabled the team to handle a higher volume of FDS documents without the need to expand team capacity, allowing the business to scale up effectively.

Additionally, automation has alleviated resource strain by handling most of the document work, freeing up engineering teams to focus on higher-value tasks such as client-specific iterations and quality enhancements.

Looking ahead to Phase 2 and beyond, the solution is expected to deliver even greater benefits. Continued development will enhance the extraction layer to handle more complex information types and new file formats. Output generation will be fine-tuned for different FDS scenarios, such as regions and business units. Solution design enhancements will include new features for better interpretability of the Gen AI solution flow and optimized costs using Azure AI services.

Custom models will be trained for accurate extraction across all document types. UI features will be enhanced to include section-specific validation and ad-hoc content generation, as well as an interactive UI for all subsections. Discovery for new use cases will explore solution design for new templates and advanced computer vision use cases, such as flowchart extraction. The solution will also be scaled for high user load and integrated with existing client systems for seamless access.