How ChatGPT revolutionizes internet search

According to media reports, Microsoft wants to invest ten billion US dollars in the company OpenAI, among other things to integrate its AI language model ChatGPT into its search engine Bing. Even if Microsoft has so far been silent about how it should be embedded in the search engine, the report has led to lively discussions among observers: Can Google now pack up or is such an integration nonsense?

First of all, there are some arguments against using language models for Internet searches. ChatGPT & Co. like to deliver fictitious stories instead of facts. And they don’t even allow users to verify the claims because they usually don’t credit their sources. Apart from that, their level of knowledge is not up to date. Language models are fed with texts up to a certain point in time, after that no more. They cannot provide any information about what happened after that.

According to OpenAI, ChatGPT’s database, for example, is up to date at the beginning of 2022. If ChatGPT knows isolated information that it must have learned later – for example that Elon Musk is the CEO of Twitter – this is only an exception to this rule . Anyone who asks the service about the Brazilian President will get the wrong answer Jair Bolsonaro.

Perplexity AI provides tailor-made answers including the sources – but the answers are not always up to date.

Search engines have been working more or less on the same principle for over twenty years: in response to a search query, they compile a list of links that they assume are relevant to the query. These links come from an index that they have created with so-called crawler programs. Such an index is much more up-to-date than the database of a language model.

Crawler, Index, Liste

Tracing services also try to answer inquiries directly where this is possible for them. To do this, they build so-called OneBoxes on the search result pages, special search results for specific questions. However, such boxes only exist for a fraction of search queries.

If a search engine could combine the advantages of both worlds – tailor-made, well-formulated answers on the one hand, topicality and verifiability on the other – it would be tantamount to a revolution in Internet searches. Several search services show that this is possible. Perplexity AI, You.com, and Neeva.com all provide ChatGPT-style, well-written responses, with which they credit their sources. The user can thus check the origin of the information.

The three search services are all immature and currently only work for English searches. Perplexity AI and You.com also operate their answer engines separately from their normal search results pages. Perplexity AI only presents free text answers and no classic link lists at all. In our tests, the search service also provided some outdated information, for example about the current President of Brazil. You.com is a relatively new, “classic” search engine. In addition, YouChat was launched in December, which generates response texts using a language model. It runs as an alternative to the traditional search.

NeevaAI writes custom text that the search engine displays with the other content boxes on the search results page.

Index feeds language model

Most coherently, Neeva has integrated its ChatGPT-style answer engine into its search service. For many search queries, their NeevaAI function shows a box on the results page with the answer dynamically written by an AI along with the sources. For example, if you search for “Twitter Files,” you’ll get the answer: “The Twitter Files are a series of documents released by Twitter owner Elon Musk that provide insight into the company’s content moderation decisions. The documents have revealed that the CIA has been meddling in Twitter’s internal content moderation for years. They have also shown that the US military has been directly assisted by Twitter.” NeevaAI also links the sources CBS News, MSN.com, Fox News and NY Post.

The service automatically decides which questions Neeva answers with free text and where to place the NeevaAI text, just like for other content that it integrates into the search results pages – images, videos and Wikipedia boxes. For example, if you type “Eiffel Tower” in the search field, you first get a link to the tower’s official website, then a map and only then the AI ​​answer. Neeva puts an excerpt from the Wikipedia entry about the Eiffel Tower to the right.

Unlike many other search services that only return results from Google or Bing, Neeva operates its own search index. Apparently, this own index is one of the bases for integrating NeevaAI so seamlessly into the search. Neeva uses several language models to answer a query, as Sridhar Ramaswamy, CEO of the service, explained in an interview with c’t. After entering the search query, a language model first determines the user’s intent and supplements the query with synonyms, spelled words and related expressions. In his blog, Neeva describes the customization of the request and NeevaAI in detail and offers feedback options if you find errors.

After NeevaAI has analyzed and rewritten the user’s request, it retrieves the content of the most relevant pages and constructs an answer from it – usually a paragraph long. The response generated by the AI ​​usually loads a little slower than the rest of the search results page – but mostly much faster than the currently heavily overloaded ChatGPT homepage. NeevaAI is currently only available for users with a US account – if you want to try NeevaAI in this country, you can simply switch your account accordingly in the settings.

Ad-free and subscription-financed

The fact that Neeva operates such a sophisticated answering engine is all the more astonishing given that Neeva, as a young search engine start-up, only has 50 employees. Neeva should remain permanently ad-free and free of charge in a basic version including NeevaAI. The company wants to make money with paid accounts. These should also search data stored in cloud services, and services from third-party providers are also included, such as Bitdefender VPN and the password manager Dashlane.

Google is certainly one of the companies with the greatest expertise in AI. The industry leader is pursuing several approaches to integrating AI into its search. The BERT language model (Bidirectional Encoder Representations of Transformers), for example, which Neeva uses to preprocess search queries, was originally developed by Google and also preprocesses queries there. And Google also has a chatbot: LaMDA – Language Model for Dialogue Applications. However, Google has not yet shown such a coherent integration of the language models into the search engine as Neeva.

Google is already under pressure in its central business area, the search. Young people are less and less likely to search on Google and instead use TikTok or Instagram. This was the result of an internal study, as a Google manager explained. And now there is also competition from other search services that can present a fully formulated answer to the search query instead of a list of links: 2023 could be the year in which search engines change more than they have for a long time.

(Bild:

c’t 4/2023

)

Photovoltaics are booming and you can’t reach an installer? In the c’t 4/2024 issue, we lend a hand and screw a PV system onto the roof in a self-experiment. c’t explains which steps you can and may carry out yourself and when the specialist company has to do it. You will get to know the legal and financial framework and learn how the PV components work. We’ve also written a practical guide to encrypting data, tried cracking USB storage with built-in encryption, and tested efficient Ryzen 7000 CPUs. You can read all this and much more in c’t 4/2023!

(jo)

To home page