AI in Public Procurement

Price research in public procurement involves consulting large public databases to ensure that the prices charged are fair and transparent. However, the process presented challenges that demanded a more modern and efficient approach.

• Access to large databases: Consulting extensive public databases, such as the PNCP, required a significant amount of manual work to locate and cross-reference relevant information.

• Limitations of textual search: Text-based searches had difficulty identifying relevant quotes for complex items and services, whose descriptions vary between suppliers and databases.

• Manual and time-consuming process: Price research consumed weeks of analysts' work, impacting the agility of acquisitions and the sector's response capacity.

To address these challenges, the team developed an intelligent price search tool, combining AI techniques with automated data collection. The approach was structured in phases that ensured precision and scalability.

Genius in action: Intelligent price search with AI

Discovery: Mapping sources and automated data collection

The first phase focused on understanding the available public databases and building the collection mechanisms. Web scraping techniques were used to access and extract data from various sources, such as the PNCP, creating a structured repository of quotes.

• Automated web scraping: Data collection from multiple public databases, eliminating the need for manual queries in each source.

• Processing with LLMs: Application of Large Language Models to remove noise from the collected data and improve dataset quality before the search stage.

Design: Semantic architecture for search accuracy

With the data structured, the design phase defined the solution's architecture with a focus on semantic search — an approach that goes beyond textual matching and understands the meaning of queries. Technical decisions prioritized efficiency and cost reduction.

• Semantic search with PG Vector: Replacement of textual search with local vector search, capable of finding relevant quotes even when item descriptions vary across databases.

• Vertex AI with Gemini 2.5 Flashlight: A choice that resulted in significant savings compared to other solutions, such as Elastic Search or GPT.

Build: Tool construction and curation module

The build phase integrated all components into a cohesive tool, including a curation module that allows the user to refine the dataset for training price prediction models.

• Curation module: Allows the analyst to refine the collected data, adjusting the dataset that feeds the predictive models.

• AI and Cloud-based architecture: The approach with Gemini allowed developers to connect ready-made functionalities instead of creating them from scratch, accelerating prototype development.

The intelligent search tool transformed the dynamics of public procurement. The price research process, which previously consumed weeks of manual work by analysts, now takes hours to complete for simple items. Semantic search brought a leap in accuracy, identifying relevant quotes that were previously overlooked in traditional textual searches.

The choice of technologies such as Vertex AI and local PG Vector ensured a financially viable solution, with significantly lower costs compared to alternatives like Elastic Search or GPT.

Furthermore, the AI and Cloud-based architecture accelerated prototype development, allowing the team to connect ready-made functionalities instead of building them from scratch.