A hard dollar limit per scrape job with model_instance (ChatOpenAI base_url + default_headers)

[domondi1]

A ScrapeGraphAI run can be one call or many: SmartScraperMultiGraph over a list of URLs, large pages split into chunks, merge steps. If you run scrapes as jobs (per customer, per batch) and want a hard dollar ceiling per job, the model_instance option is enough. Pass a LangChain ChatOpenAI pointed at a local gateway, with the job id and budget as default headers. I maintain the gateway (Inferrail, open source).

pip install scrapegraphai langchain-openai inferrail
export OPENAI_API_KEY=sk-...
inferrail serve --quickstart --app-mode      # http://127.0.0.1:8000

from langchain_openai import ChatOpenAI
from scrapegraphai.graphs import SmartScraperMultiGraph

def config_for_job(job_id: str, budget_usd: str) -> dict:
    llm = ChatOpenAI(
        model="gpt-4o-mini",
        base_url="http://127.0.0.1:8000/v1",
        api_key="unused",                       # the real key stays in the gateway
        max_tokens=800,
        default_headers={
            "X-Inferrail-Attribute-Work-Id": job_id,   # one id per scrape job
            "X-Inferrail-Budget-Usd": budget_usd,      # its dollar ceiling
        },
    )
    return {"llm": {"model_instance": llm, "model_tokens": 128000}}

graph = SmartScraperMultiGraph(
    prompt="List every product name and price.",
    source=urls,
    config=config_for_job("catalog-batch-17", "1.50"),
)
result = graph.run()

Every LLM call in that job carries the same id, so they share the $1.50. A call that would go past it gets HTTP 402 before it reaches OpenAI. Then inferrail work catalog-batch-17 prints calls, tokens and dollars for the job, and inferrail report --by work_id lists all jobs.

What I checked (scrapegraphai 2.3.0, langchain-openai 1.6.7, inferrail 0.4.15, SmartScraperGraph on raw HTML against a stand-in OpenAI endpoint with fixed token counts):

• The call was recorded under the job id with tokens and cost; the X-Inferrail-* headers are not forwarded upstream.
• Budget too small: graph.run() raised openai.APIStatusError 402 (INFERRAIL_E010, with the budget and the estimate in the message) after one attempt, and nothing reached the upstream. Catch it if you want partial results instead.

Caveats: a dollar budget needs a price for the model (most OpenAI and Anthropic models are built in, inferrail models; an unpriced model is refused, not guessed), and the gateway stores usage and cost per call, never page content or prompts. I only tested the single-page graph; the multi graph uses the same model_instance.