Hel.fi main search is the site search for the core section of the hel.fi website (www.hel.fi domain). The search improves the discoverability of City of Helsinki services and other information important to residents.
The service is intended for anyone interested in the City of Helsinki’s services and activities. The search can be used to find information on web pages using natural language and questions. An AI-based language model adds an assessment describing the meaning of the hel.fi content and the user’s search phrase. Search results are ordered by comparing the similarity of the meanings.
Contact information
Department
City Executive Office, Digitalisation Department
Yhteydenottotiimi tiedusteluja varten
Drupal Solutions Office (Aleksi Salonen, Mari Pietarila)
External suppliers
None.
Contact email address
Detailed system information
Here you can check what information the system uses, how it works and how its use is monitored.
AI training data
OpenAI’s text-embedding-3-large vectorisation model, procured as an off-the-shelf product, is used to analyse the content of the Hel.fi website and users’ search phrases. The model’s training data is the supplier’s trade secret. Content indexed for the search and search phrases entered by users are not used as training data.
The dataset contains the public hel.fi website
The dataset shown to the search user contains public web pages on the hel.fi website and search promotions manually created by City employees. No content generated by the language model is produced for the user.
System architecture
The natural language analysis performed by Hel.fi main search is based on OpenAI’s text-embedding-3-large data vectorisation model, which is procured through the Microsoft Azure cloud service with usage-based pricing. The service is located in the EU, in the Sweden Central Data Zone Standard region. Embedding data concerning Hel.fi content is stored in the Hel.fi Elastic search index, where the user’s search phrase and the search dataset are also compared and ordered. The service uses a language model to understand the content of the hel.fi website and user questions expressed in natural language. Search results are ordered according to their relevance to each other.
Performance
The development of the service’s use and quality is assessed using analytics and user satisfaction surveys. Quality indicators include, among other things, ease of use and the proportion of users who found what they were looking for.
The service is available in three languages (Finnish, Swedish and English), but it also supports search phrases in other languages in accordance with the capabilities of the language model used. Search phrases support natural language, users’ spelling errors and imprecise search terms, making information easier to find for all users.
Web analytics connected to the search are used to monitor frequently asked topics and search phrases that produce poor responses. Based on the analytics, improvements are made to search promotions and, where necessary, to the technical optimisation of the search.
The service never asks for personal data that would enable a person to be identified. In the search analytics, personal data related to users (IP address, etc.) is automatically anonymised. The service does not include two-way communication.