Implement a 3-Position Web Search Switch for Speed vs. Accuracy Control

The current web search toggle is unreliable because the system tries to guess whether a search is needed to optimize for speed. This often results in outdated answers from the training data (pre-April 2024) even when the user expects current information, or unnecessary delays when quick answers are sufficient.

Users report that they often have to explicitly type "search the web" or add keywords like "latest" to force a search, which defeats the purpose of a simple toggle. The current behavior creates a "guessing game" where users never know if the answer is real-time or historical until it is too late.

I propose replacing the binary toggle with a 3-POSITION SWITCH to give users full control:

OPTION 1: OFFLINE / TRAINING ONLY (SPEED MODE)

Disables web search completely.

Uses only the model's internal knowledge (fastest response time).

Ideal for creative writing, coding, summarizing old texts, or when current events are irrelevant.

OPTION 2: AUTO (SMART MODE)

The AI decides if a search is needed based on the query.

This is the current intended behavior but needs to be fixed to be deterministic and reliable.

Ideal for general questions where the user is unsure if the info is current.

OPTION 3: FORCE WEB SEARCH (ACCURACY MODE)

Always performs a live web search, ignoring internal training data for facts.

Guarantees up-to-date information (slower response time due to search latency).

Essential for news, stock prices, weather, sports scores, and current events.

WHY THIS MATTERS: This solves the conflict between speed and accuracy. Users who need instant answers can choose "Offline," while those needing verified current data can choose "Force Search." It eliminates the frustration of guessing and makes the tool predictable. Currently, the inconsistency drives users to competitors like ChatGPT who handle this more transparently.

TECHNICAL FEASIBILITY: This is a simple UI and logic change. It involves passing a specific flag to the model router: "Ignore internal knowledge for facts" (Force) vs. "Use internal knowledge only" (Offline). No new backend infrastructure is required, just a clearer user interface and stricter logic for the "Auto" mode.

Nearly a year later and this appears to still be an issue.

Web search is repeatedly NOT triggered despite toggle on AND despite some phrase like "Search the web" in the prompt. Really! -- BTW, the toggle is misleading! It does NOT guarantee web ssearch is actually triggered. And now Lumo+ told me today that even "Search the web" is not sufficient sometimes that something like "latest, recent etc." is also needed in the prompt. Are you serious? Sorry, you might read frustration in my words. Yes, take it as a prove that I have had this issue too often. The Lumo web search feature is a MESS and unreliably triggered.

New user of Proton and Lumo. Ran into the same thing from Lumo after toggling web search on a subject I knew may take a little more digging. Got results, they were not what I was looking for, so asked Lumo again. Got the 2024 dataset comment. And Lumo will also more or less slightly change the answer and just repeat over and over again if you keep prying. I am new to finally getting real business services like this for myself now that my demand has risen to the point it's essential. A.I. assistants I'm especially new to using on a consistent basis. So I can't say the repeats aren't on myself for not being overly experienced in prompting A.I.

Just an option in settings to default web search to on would be awesome and simple.

Here is what Lumo told me, and it works.

With the web search enabled are you able to switch back-and-forth between models and search?

Absolutely!
How it works under the hood
Model routing – Lumo runs several specialized language models (e.g., Mistral‑Small, OpenHands, OLMO 2, Nemo). When a request comes in, an internal router decides which model is best suited for the task—coding, reasoning, summarisation, etc. That decision happens automatically and can change from one turn to the next.

Web‑search integration – If the router determines that the current query needs up‑to‑date information (news, weather, stock prices, recent product releases, anything after our April 2024 knowledge cut‑off, etc.), it triggers the web_search tool. The search result is then fed back into the chosen model so the final answer combines the model’s reasoning with fresh data.

Back‑and‑forth flow
You ask a question.
The router picks a model.
If the model flags a need for current info, it calls the web‑search tool.
The search result returns, the model re‑processes the combined context, and replies.
On the next user turn, the router may pick the same model again—or a different one—depending on the new request.

So yes, with web‑search enabled Lumo can seamlessly hop between different models and the search tool as many times as the conversation demands, all while keeping the experience smooth for you.

69 Votes over 3 topics. This is the single largest issue for Lumo.

If this is not something that can be corrected quickly, then provide us a workaround, or documentation regarding how we can reliably utilize the function reliably.

I believe this is covered in Lumo - Consistent Web Search.

I have been running into this problem significantly recently, where I click "Web Search (WS)" and ask my question, but Lumo does not search the web for an answer. Instead I have to ask again "Search Web for _______", and I get an often times, vastly different answer. It seems to me that there should be a way to guarantee that Lumo (at least) considers a WS when the button is checked. I understand not all queries require WS, but one of the main points of using AI is to synthesize information, and rendering opinions or information about current events without utilizing web search is leading to inaccuracies and dated information.

As a side note, I believe there should be a way to permanently select WS so that it does not have to be clicked every time.

Thanks for the hard work and attention to user feedback!

This makes me jump back to ChatGPT even though i really don't want to. It's just extremely inconsistent and bad.

As a paying customer, I frequently rely on the web search functionality in Lumo to enhance my experience and get valuable information. Currently, the web search option resets when I close and reopen the tab, which is inconvenient.

I would like to suggest that the web search option be made persistent and saveable to my account settings when I'm signed in. This way, I won't have to enable it every time I use Lumo, making my workflow more efficient.