How APA wants to deliver a “better Google” » Leadersnet

Technology department in Austrian Press Agency APA was invited to attend a technical lecture on October 8 on the topic “How AI-powered technologies and bots improve search results and user experience.” Challenges and opportunities of semantic search were also discussed and a specially trained AI solution from APA Group – a “Google competitor” from Austria – was presented.

background APA Technical Talks It is a development that is now part of daily work in many companies – the use of ChatGPT or similar large language models (LLM). Anyone who works with AI assistants is probably familiar with this: the same prompt yields different results, and there is no trace of transparency or repeatability. This makes scaling up to different applications as a business case, for example in e-commerce or media use, a challenge. APA-Tech wants to provide suitable solutions for this. The technical talk also included information about the strategy being followed in the field of semantic search and related technologies and how to develop and use them within the company. It was also presented what was learned in the first client project and what the news agency will provide to its clients in the future.

Advantages of semantic search

Semantic search makes it possible to search not only for keywords, but also for meanings and associations, and provides relevant results even if the exact words do not appear in the search query. It is considered a turning point in the way people search, find and process information. However, semantic search can be affected, he explained Andreas MuchkaAPA's digital director, using your content to train content focus in the basic form. This is also the basis for the reasonable use of the so-called RAG (Retrieval Augmented Generation) concept. By doing so, the linguistic model is provided with its own qualitative database that is as accurate as possible, limiting its options for “creativity.” The head of digital affairs stressed that “the model does not require Google searches or hallucinations, but rather provides fact-oriented results.”

Client project – from prototype to business case

The first major customer project has been reported Sophie Hlinka,Product Owner APA-Tech, Mika gerus From the archive service provider and content marketer DIZ Munich, where the archive is also located South German newspaper It is turned on. “Archive searches have to become more efficient in order to save time on fact-checking,” said Giros. They wanted to offer the editorial team and archive clients the best possible search experience based on their own data.

APA, which has been working on semantic search for years, was the logical choice for DIZ for a joint proof-of-concept project. All APA Group's data and content management services are based on the “PowerSearch” full-text database technology developed by APA-Tech, the functionality of which has been extended to include a vector database. “All services today rely on very specific queries, including semantic searches that barely capture the intent behind the query,” he explained. Klemens Prirowskimanaging director of information technology at APA.

“In contrast to purely technical service providers, in our model, we have access to the daily intellectual work, i.e. the content of the editorial team of APA and APA-PictureDesk, and provide clients with artificial intelligence services trained by APA in Austria,” so Mauchka.

For DIZ, APA developed an interface with different testing functions, such as semantic search, Q&A (question/answer) or chatbot application. Since the standard model did not provide satisfactory results, it was decided to train the client-specific model and build an embedded model for semantic search. For data protection reasons, the language model chosen was not ChatGPT, but a local setting using the European Mistral model. ChatGPT has spoiled users in terms of performance and costs. “Our goal is to achieve similarly good results using smaller models,” he says. Hlinka.

Google is better

“We're not Google, but we're still a big company: there are 1.4 billion documents in the APA PowerSearch database, and millions of data records are added month after month,” he explains. Prirovsky Unique selling point. In addition, well-maintained metadata will make the database a verified database. “We have the technological foundation and can easily adapt and scale models for customers at any time,” he promises Prirovsky.

In the APA model, there is no contradiction between artificial intelligence and truth-oriented information;

www.apa.at