AI simplifies information retrieval
“Participating in the testing of an AI-assisted chat interface was a new and interesting experience for us. We learned that in addition to training the AI, users also need to be taught how to formulate questions so that the AI can find a reliable answer from a large amount of data.”
Sari Suomela, Chief of Records Management, The City of Turku
The need
The City of Turku needed to develop a user-friendly solution to make the information in the city’s information management system, JoutseNet, more easily accessible. Many organizations face a similar challenge: there is a vast amount of data, but finding the exact information or document is challenging.
The solution naturally had to ensure that the data remained in Europe in accordance with GDPR and that the data did not leak outside the organization. While services like ChatGPT can use the information fed into them for further model training, organizations can build systems based on large language models where this does not happen.
The organization
Turku is the largest city in Southwest Finland, the regional center of Finland Proper, and the sixth-largest municipality in Finland in terms of population.
The city develops and offers comprehensive digital services to the citizens, taking care of its statutory tasks and obligations.
The Solution
ATR Soft created a Proof of Concept (PoC) level solution for Turku, a chat interface that can be asked anything about the city’s public decisions.
In the solution, we utilized Azure AI Search and Azure OpenAI language models and vectorized JoutseNet documents, such as decision minutes. Vectorization transforms texts into numerical representations so that their context and meanings remain analyzable. We imported these text vectors into the Azure AI Search index, from which the solution can search texts and documents using vector search. The search essentially compares the meanings of search phrases rather than just looking for exact matching words.
The user can write their question to the chat, after which the question’s vector is compared to the documents in the search index. The suitable documents and the original question are then given to a large language model, such as the Azure OpenAI GPT-3.5 model, which forms an answer to the user’s question using the information in the documents.
The project demonstrated that AI can significantly improve the use of the city’s information management system. The results achieved and lessons learned in the PoC phase provide a solid foundation for further development, which can create an even more efficient and user-friendly solution for the city’s needs.
Technologies
Azure OpenAI: Embedding models can calculate vectors, i.e., numerical representations, for texts. Large language models (LLMs), such as GPT-3.5, can generate answers. By modifying prompts, it can be ensured that the answer is based on the provided background information, thus minimizing the model’s hallucination, i.e., the use of incorrect information.
Azure AI Search: By using a search index optimized for vector search, relevant information can be found and retrieved more accurately because the search is based on meaning and context, not just exact keywords.