Voice-powered web search

Asking "what's the weather like tomorrow" to an AI that never leaves its own head is like inviting it to guess. We've integrated real web search into the conversation, and learned that not all ways of searching are equal.

Answering accurately, or answering quickly

A language model can answer almost anything, even when it doesn't know: this is the classic pitfall of conversational AI, a confident answer that is sometimes just a guess dressed up as a fact. For a question about the day's weather, breaking news, or travel time, there's no honest shortcut: you have to go find the information, not guess it from what the model learned on a given day. We already had a first web search mechanism integrated into the conversation, but it showed its limits on multi-step questions (a multi-segment journey, a duration to be reassembled from several scattered pieces of information). "How long does it take to get to Lyon via Chambéry" requires searching for two journeys, adding them up, then checking if the itinerary makes sense: the first mechanism often treated this as a single flat question, and the answer suffered. The kind of question that, once again, encourages guessing rather than proper searching.

Changing to a real engine

So we switched to a web search engine natively designed for this kind of use, rather than an extension grafted onto a general-purpose language model after the fact. The difference is especially noticeable on compound questions: a multi-step travel request is no longer treated as an isolated question, but as a real search context, with several sources cross-referenced before answering. The model used for this task remains configurable server-side, without depending on a fixed ID in the code. Useful for changing versions as soon as a better one appears, without rewriting the circuit that calls it.

Sources, not assertions

Every answer resulting from a web search comes with the sources that fed it. This isn't a cosmetic detail: it changes the nature of the answer. XNeuronal no longer simply says "it will rain tomorrow," it can tell you where that information comes from, and you retain the ability to verify it yourself if the subject warrants it. A useful answer can also become a memo card accessible later, sources included, rather than getting lost in the conversation thread once the next sentence is spoken. This is the mechanism you'll find today in XNeuronal's web search.

The boundary between remembering and searching

This project also clarified a boundary that should never be blurred: XNeuronal's memory retains what you tell it, web search goes to find what you don't know yet. The two should never substitute for each other: a fact you mentioned should not trigger an unnecessary external search, and a question about the day's news should never be guessed from an old, poorly adjusted memory, even a recent one, even if formulated with confidence. It's a distinction that seems obvious on paper, but one that must be maintained with each turn of conversation, question after question, without ever letting it fade.

The compromise we accept: a web search structurally takes longer than an answer drawn from immediate memory, because it actually has to consult the outside world before answering. We prefer this accepted delay to an instant answer that could have been wrong: better one or two more seconds than an invented weather forecast delivered with aplomb.