Porto do Açu cuts document processing time by 99.5% with AI

Porto do Açu faced significant challenges in searching for information within EIA documents, which include data in different formats, such as .txt, .pdf, .xlsx, .docx, .tiff, .json, .jpeg and .png. The process of searching and extracting information was time-consuming, which negatively impacted response times for regulatory queries. Furthermore, keeping the virtual library up to date and managing information efficiently required a robust solution that could deal with these challenges in a scalable and accurate way.

MadeinWeb developed a Generative AI solution supported by the AWS infrastructure, creating a Generative Chat that optimizes the environmental licensing process. This solution allows quick and accurate queries to a virtual library made up of technical manuals, standards and EIA data. The solution is based on a Recovery Augmented Generation (RAG) system, which searches a vast set of data provided by Porto do Açu, facilitating access to relevant information with precision.

Amazon Bedrock: We use Bedrock to access the Claude3-Haiku text generation model and the Titan Embedding V1 Embedding Model. The choice for Claude3-Haiku is due to the low complexity of the task, which is to answer questions based on a context.

1. Interaction via generative chat: The Generative Chat was developed to directly access the virtual library of EIA documents and provide quick and assertive responses to users.

2. Document parsing: The solution included advanced parsing techniques to extract data from documents in different formats:

   • PDFs and DOCX: They were processed using the Llama-Index to ensure accurate extraction of text and tables.

   • Images: We use Bedrock's Claude3-Haiku to convert images (.jpeg, .png) into text, ensuring the interpretation of graphs and tables.

   • XLSX Tables: Each spreadsheet tab has been converted to format HTML, facilitating interpretation by language models such as Claude Instant and Claude2, allowing the content of tables to be consulted with high precision.

3. Document segmentation: Text documents were divided into blocks of 900 characters with overlapping 200 characters to improve performance and processing speed. For table data, the solution generated documented descriptions, facilitating accurate retrieval of this information.

4. Metadata and filters: The system used metadata to organize and manage documents:

   • File name: Used for management in the administrative interface.

   • Relevant EIA: Chat asks the user which EIA is being consulted, filtering the information based on this metadata.

   • Data source: If the information comes from a text or table, the solution retrieves the complete context of the tables stored in S3, ensuring that the model responds accurately using complete data.