Web Search
How your assistant looks things up online — the two modes (managed vs bring your own), the four built-in providers, the fallback chain, and how it's billed.
Overview
Web search lets your assistant pull live information from the internet to answer questions, verify claims, and ground its responses in current sources. Whenever a conversation, scheduled task, or skill needs information that isn't in your workspace, the assistant can call out to a search provider, read the results, and cite the sources it used.
You control which provider runs that search and how it's billed. You can use the native search capability bundled with your inference provider, or connect your own API key for a dedicated search engine like Perplexity, Brave Search, or Tavily.
Managed vs Bring Your Own
Web search runs in one of two modes.
| Mode | What it does | Who pays |
|---|---|---|
| Managed | The assistant uses the native search capability built into your selected inference provider. No separate search vendor receives the query. | Billed through Vellum credits as part of inference. |
| BYOK | The assistant calls a dedicated search provider with the API key you connect. You pick the provider. | Billed directly by the provider under your account. |
Most workspaces start in managed mode and stay there. Bring Your Own makes sense when you want a specific search engine's results, when you already have a Perplexity / Brave / Tavily account, or when you want search costs separated from inference costs.
Providers
Four providers ship with every workspace.
| Provider | Kind | API key prefix | Notes |
|---|---|---|---|
| Provider Native | Managed | n/a | Hands the search to the inference provider. Quality and freshness follow whichever LLM you have selected. |
| Perplexity | BYOK | pplx-... | Synthesized answers with inline citations. Default first choice in the fallback chain. Pulls from the Perplexity Search API. |
| Brave Search | BYOK | BSA... | Independent index, no tracking, supports freshness filtering. Good default for privacy-leaning workspaces. |
| Tavily | BYOK | tvly-... | Search API designed for AI agents. Returns scored results with extracted content blocks. |
For each BYOK provider you connect, your assistant stores the key locally and uses it for every web search until you disconnect or switch providers. Provider privacy policies: Perplexity, Brave, Tavily.
Choosing a provider
Side-by-side comparison across the dimensions that usually matter.
| Dimension | Provider Native | Perplexity | Brave | Tavily |
|---|---|---|---|---|
| Setup | None. Works out of the box. | API key required | API key required | API key required |
| Billing | Bundled with inference (Vellum credits) | Direct to Perplexity | Direct to Brave | Direct to Tavily |
| Result style | Depends on the LLM you selected | Synthesized answer with inline citations | Raw ranked results (title, URL, snippet) | Scored results with extracted content blocks |
| Index source | Whatever the LLM provider uses | Aggregates multiple search engines | Independent crawl, not Google or Bing | Aggregates multiple search engines |
| Freshness control | Not exposed | Not exposed | Supported (day / week / month / year) | Supported |
| Privacy | Subject to LLM provider terms | Standard SaaS | No query tracking | Standard SaaS |
| Agent ergonomics | Hidden behind the LLM | Pre-synthesized, agent can quote directly | General-purpose search shape | Built for agents (scores, content blocks, raw content) |
| Best for | Default setups, low overhead | Cited answers in chat | Privacy-leaning workspaces | Agentic workflows with extraction |
If you're unsure, start on Provider Native and switch only when you hit a concrete reason to.
Configuring web search
Web search is configured in Settings → AI → Web Search.
- Choose a mode: Managed or Bring Your Own.
- In Managed mode, no further setup is required. Searches use the native capability of your selected inference provider.
- In Bring Your Own mode, choose a provider (Perplexity, Brave, or Tavily) and paste your API key into the field next to it. Keys are stored locally on the device that runs your assistant.
- Optionally connect keys for more than one BYOK provider. The extras become fallbacks if the primary provider fails or runs out of quota. See Fallback behavior below.
Changing the provider takes effect immediately for the next search; in-flight searches finish on the previously configured provider.
Fallback behavior
In Bring Your Own mode, your assistant tries the provider you selected first. If that provider has no key configured or its request fails with a retryable error, the daemon walks the fallback chain in this order:
- Perplexity
- Brave Search
- Tavily
The chain skips any provider that doesn't have a key connected. If none of them have keys, the search returns an error and the assistant tells you it couldn't reach the web. Connecting more than one BYOK key is the simplest way to keep search resilient when an upstream provider has an outage.
Billing
Web search is one of four chargeable categories in the Vellum platform, alongside LLM inference, image generation, and paid third-party APIs.
- In Managed mode, search cost is bundled into the inference call that triggered it. You pay Vellum credits the same way you pay for any other LLM request.
- In Bring Your Ownmode, the search provider bills you directly under your account with them. Vellum doesn't mark up or proxy these requests.
For pricing details and credit denominations, see the pricing page.
