misc_action.llm_query sends a prompt directly to an LLM and stores the structured response—no browser state, axtree, or screenshot is involved.
Overview
- Use when: You need to transform, classify, or reason over data already in memory (e.g., reformat an extracted date, classify extracted text, compute a derived value).
- Execution: The action sends
prompt_instructions to the configured LLM, parses the response according to output_format, and stores results in output_data. Specified fields are optionally promoted to generated_variables.
- No browser access: Unlike
extraction_action.llm, this action does not read the current page. Reference earlier extracted values via {variable_name[index]} in your prompt.
Properties
| Property | Type | Default | Description |
|---|
output_format | dict | Required | Expected output structure as a type-annotated dict |
prompt_instructions | str | Required | The full prompt to send to the LLM |
output_variable_names | list[str] | None | None | Promote response fields to generated_variables |
llm_provider | "gemini" | "anthropic" | "openai" | "gemini" | LLM provider |
llm_model_name | str | "gemini-2.5-flash" | Model name |
JSON Example
{
"type": "action_node",
"misc_action": {
"llm_query": {
"output_format": {
"category": "str",
"confidence": "str"
},
"prompt_instructions": "Classify the following support ticket as 'billing', 'technical', or 'other'. Ticket: {ticket_text[0]}",
"output_variable_names": ["category"]
}
}
}
{
"output_format": {
"summary": "str",
"status": "str",
"items": "List[str]"
}
}
Only str and List[str] are supported types in output_format.
Using Variables in Prompts
Reference extracted values from earlier nodes using {variable_name[index]}:
{
"misc_action": {
"llm_query": {
"output_format": { "normalized_date": "str" },
"prompt_instructions": "Reformat this date to ISO 8601 (YYYY-MM-DD): {raw_date[0]}",
"output_variable_names": ["normalized_date"]
}
}
}
{normalized_date[0]} is available in subsequent nodes.
Storing Output as Variables
Set output_variable_names to promote response fields into generated_variables:
{
"misc_action": {
"llm_query": {
"output_format": {
"next_url": "str",
"has_more_pages": "str"
},
"prompt_instructions": "Given these pagination links: {page_links[0]}, return the next page URL and whether more pages exist ('true'/'false').",
"output_variable_names": ["next_url", "has_more_pages"]
}
}
}
Every key in output_variable_names must appear in output_format. Validation fails at schema load time if a key is missing.
Choosing a Provider and Model
| Provider | Models | Notes |
|---|
gemini | gemini-2.5-flash, gemini-2.5-pro | Default; cost-effective |
anthropic | claude-sonnet-4-6, claude-haiku-4-5-20251001 | Strong reasoning |
openai | gpt-4o, gpt-4o-mini | OpenAI models |
llm_provider and llm_model_name are omitted, the task-level defaults are used.
| misc_action.llm_query | extraction_action.llm |
|---|
| Browser context | None — prompt only | axtree and/or screenshot of current page |
| Typical use | Transform / classify in-memory data | Extract data from the page the browser is on |
| Input source | Variables embedded in the prompt | Live page content |
| Output destination | output_data + optional generated_variables | output_data + optional generated_variables |
Use misc_action.llm_query after an extraction step to post-process or validate extracted data without consuming a browser round-trip.
