Aevum Cadence AI

Web search

Web search lets the assistant pull live information from the internet to answer questions, verify claims, and cite sources. When a conversation, scheduled job, or skill needs something that is not in the workspace, it can call a search provider, read the results, and use them in the reply.

In 0.1 there is no Aevum-metered Cadence search. You bring your own key for Brave, Perplexity, Tavily, or Firecrawl, or you use a model provider that already has native search. Web Fetch labeled Cadence in Settings is the local builtin fetch, not a hosted search SKU.

Providers

ProviderKindNotes
Provider native Model Uses the selected LLM's own search when that model has it. Quality follows the model. If it has none, the assistant tries any connected BYOK search key.
Brave Search BYOK Independent index, no query tracking. Good default for privacy-leaning workspaces.
Perplexity BYOK Synthesized answers with inline citations. Default first choice in the BYOK fallback chain.
Tavily BYOK Search API shaped for agents: scored results with extracted content blocks.
Firecrawl BYOK Search plus full-page scrape. Clean markdown, including JavaScript-rendered pages. One key can also power web fetch.

Provider privacy policies: Brave, Perplexity, Tavily, Firecrawl. Queries leave your machine and reach that vendor.

Choosing a provider

There is no Google search provider. Do not expect a Cadence-billed search balance in 0.1.

Configuring

Settings → Models & Services → Web Search. Pick a provider, paste the API key. Keys stay in the local credential store. CLI:

assistant keys set brave BSA...
assistant config set services.web-search.provider brave

You can connect more than one BYOK key. Extras become fallbacks if the primary fails or runs out of quota.

Fallback

On a BYOK provider, the assistant tries the one you selected first. If that key is missing or the request fails, it walks this chain and skips any provider without a key:

  1. Perplexity
  2. Brave Search
  3. Tavily
  4. Firecrawl

If none have keys (and the model has no native search), the search returns an error and the assistant tells you it could not reach the web. Connecting more than one key is the simplest way to stay resilient through an upstream outage.

Billing

BYOK vendors bill you directly. Cadence does not mark up or proxy those requests, and it does not sell search credits in 0.1. Model native search, when it exists, is billed by that model provider the same way as any other inference call.

Related: Model profiles, Features, Browser.