Web search is off by default. When enabled for a message, LADLE searches, retrieves the most relevant sources, and returns an answer with inline citations that link to the original URLs. This page documents exactly how the retrieval works, what's cited, and what to verify.
How to enable search
Three ways to turn search on for a message. Each has an equivalent visible affordance in the composer.
What the search step does
When search is triggered, LADLE queries the web, pulls up to 8 top results (filtered for relevance, freshness, and authority), extracts the passages most relevant to your question, and passes them to the model along with your prompt. The model composes an answer that cites those passages. Search adds latency of 1-3 seconds compared to a no-search response.
How citations render
Cited claims in the answer have numbered superscripts inline: `The GPAI provisions took effect on 2 August 2025[1].` The right rail (or, on mobile, the footer of the answer) lists sources in order — number, title, domain, and a link out. Click a citation superscript to jump to the source in the rail; click the source in the rail to open the URL in a new tab.
What gets cited, and what doesn't
Any specific factual claim retrieved from search is cited. Explanatory scaffolding around those claims — connective sentences, background you might already know, LADLE's own paraphrasing — isn't. If an answer isn't cited, it's from the model's general knowledge, not search. The rail's `NO SEARCH` state indicates the entire response came from general knowledge; treat it accordingly.
Verifying sources yourself
Every citation links out. Click through and check the source says what LADLE claims. Where a source's date matters (news, regulations, prices), the date is shown inline with the citation. If you can't verify a claim from the source, that's a bug — please report via help@ladle.chat with the chat link.
Search inside a Project
Search operates the same way in a project chat — the project's files aren't part of the web search. If a project file contradicts a web source, LADLE surfaces both and defers to you on which to trust. Project files are treated as attached context; web results are treated as fresh retrieval.
What search costs in usage
Retrieved passages consume input tokens (see /docs/usage-limits). A typical 8-source search adds ~4,000 input tokens beyond your prompt. Included in every plan tier at the same limits as regular messages.