An AI node calls a language model with a prompt and hands the result to the rest of your workflow. Use it to classify, extract, summarize, draft, or decide. AI nodes bill in AI Credits by actual token cost: bigger models and longer outputs cost more per call. See billing .
#Prompt
- Prompt (required), the instruction for this step. Reference earlier data with
{{node_name.field}}and{{trigger.field}}. - System prompt (optional), a leading instruction that sets role, tone, or rules independent of the per-run input.
#Model
Pick the model in the builder’s model dropdown. The dropdown is always the live list: new models show up there first, so treat the list below as a snapshot, not the source of truth.
| Family | Current models | Best for |
|---|---|---|
| Claude Opus | 4.8, 4.6, 4.5 | Heavy reasoning, hard judgment calls |
| Claude Sonnet | 4.6, 4.5 | General-purpose work, good default |
| Claude Haiku | 4.5, 3.5 | Fast, cheap classification and extraction |
| GPT | 5.2, 5, 4.1 | GPT-5-class for reasoning, 4.1 for fast/cheap |
Heavier models spend AI Credits faster; reach for them only when a task needs the reasoning.
#Generation settings
| Setting | Default | Range | Notes |
|---|---|---|---|
| Temperature | 0.7 | 0–2 | Lower is more deterministic; raise it for creative drafting. |
| Max tokens | 1024 | 1–4096 | Caps the output length. |
#Output mode
- Text (default), the model returns free text, available downstream as the node’s output.
- JSON, the model returns structured data. Supply an output schema describing the fields you expect, and downstream nodes get typed suggestions in the variable picker. Use this when the next step needs to branch on or read specific fields.
#Retry policy
AI nodes support the standard retry policy, when a call fails with a retryable error, Rills retries before marking the step failed. The defaults and ranges are the same as every other retryable node; see the retry table .
#Research tools
An AI node can reach the web mid-run. Enable one or both:
fetch_url, fetch and read a specific page.web_search, search the web and read results.
Use them to enrich a lead from its company site, check whether a URL is live, or pull a fact the prompt needs. Both add fetched pages to the model’s context, which spends more tokens, so guardrails bound the cost:
- Max research pages,
1–20, default5. Caps how many pages a single run may fetch. - Allowed domains (optional), an allowlist for
fetch_url. When set, a fetch is rejected unless its hostname matches, or is a subdomain of, a listed domain. Leave it empty to allow any domain. Listingacme.comalso permitsblog.acme.comandapp.acme.com.
Research tools only do anything when the node’s output depends on fresh external data; leave them off for pure reasoning or transformation prompts.
#Workspace context
By default, every AI node is given your workspace’s business context, a snapshot of what your company does, who you serve, and how you talk, injected as a leading system block. It’s the same context you set in your workspace settings and refine during onboarding, and it means you don’t have to re-explain your business in every prompt.
Two modes:
- Workspace (default), inject the snapshotted business context.
- None, opt out. Use this for tightly-tuned prompts that must see only exactly what you wrote, no workspace framing.
The context is an immutable snapshot; AI nodes never see mutable conversation history or per-user memory.
#See also
- Workflows : the full node reference and retry policy
- Billing & Credits : how AI Credits are metered
- Runs & Versions : inspecting a run’s AI step output