AI Web Scraper: JSON Schema Extractor
Pricing
from $2.10 / 1,000 page extracted (your own llm key)s
AI web scraper with strict JSON Schema extraction: give URLs and a schema, get validated JSON per page. HTTP-first, so far cheaper than browser scrapers. Bring your own LLM key at $0.003 per page, failed pages are never charged.
Pages to extract data from. Static or server-rendered HTML pages work; pages that need JavaScript to show content are reported as 'needs-javascript'. If 'Follow links' is on, these are listing pages and detail pages are extracted instead.
JSON Schema (draft-07 style) of the object to extract, with type 'object' at the top. Property descriptions guide the model. Output is validated against it. Leave empty to use 'Fields' instead.
Used only when 'JSON Schema' is empty. One field per line: 'name', 'name: type' (string, number, integer, boolean, string[], number[]) and optionally ' - description'. Example: 'price: number - price without currency symbol'. All fields are required but nullable.
Optional plain-language hints for the model, for example 'Prices are in EUR; ignore sponsored items'. Max 2,000 characters.
Whose API your key belongs to. 'OpenAI-compatible' covers OpenAI, OpenRouter (set Base URL to https://openrouter.ai/api/v1), Groq, Together, Mistral and self-hosted servers. Ignored when billing mode is 'inclusive'.
Your LLM provider API key. Tokens are billed by your provider, not by this Actor. Stored as a secret and never logged. Empty = preview mode (no LLM call, no charge).
API base URL. Leave empty for OpenAI (https://api.openai.com/v1). For OpenRouter use https://openrouter.ai/api/v1. A custom Base URL needs an explicit 'Model'.
'fast' = cheapest small model (default: gpt-4.1-mini, claude-haiku-4-5, gemini-2.5-flash-lite). 'accurate' = stronger model (gpt-4.1, claude-sonnet-4-5, gemini-2.5-pro) for messy pages and complex schemas. Overridden by 'Model'.
Exact model name to use instead of the tier default, for example 'google/gemini-2.5-flash' on OpenRouter. Required with a custom Base URL.
'byok' (default): you use your own API key; this Actor charges the lower 'page-passthrough' event and your provider bills the tokens. 'inclusive': no key needed, LLM cost included in the higher 'page-inclusive' event; uses the fast tier only and is available only if the Actor owner enabled it.
How many times to send the validation errors back to the model when its JSON does not match the schema (0 to 3). Each retry costs tokens. Pages that still fail are not charged.
Stop after this many successfully extracted pages. Also the upper bound of what the run can charge.
Treat each URL as a listing page: find links to detail pages and extract those instead (the listing itself is not extracted or charged).
Only with 'Follow links'. CSS selector of the links (or of containers holding them), e.g. 'article.product_pod h3 a'. Empty = automatic detection of the main group of similar links.
Only with 'Follow links'. Regular expression a detail URL must match, e.g. '/catalogue/.+_\d+/'. Empty = no filter.
Hard limit of detail pages followed from each listing page.
Hard limit of all HTTP page fetches in the run (listings and detail pages), including pages that fail.
Restrict the text sent to the model to this part of the page, e.g. 'main' or '#product'. Smaller input = fewer tokens and better accuracy. Empty = whole page without scripts, navigation and footer.
Keep image links in the Markdown. Turned on automatically when the schema mentions image, photo, thumbnail, picture or logo.
Add the cleaned Markdown sent to the model to each dataset item (useful for debugging, makes items larger).
Page Markdown longer than this is cut (roughly 4 characters per token). Lower = cheaper. Inclusive mode caps this at 40,000.
Pages processed at the same time. Lower it if your LLM provider rate-limits you.
Timeout for each page download.
Proxy for downloading pages. Default is Apify datacenter proxy. Sites that block datacenter IPs may need the residential fallback below.
{ "useApifyProxy": true}