Short answer. Use
plugsky-parse, a vision-capable model with a 262,144-token context and up to 16,384 output tokens. Send each page as an image (or extract the text first), provide a JSON schema, and validate the response before writing it to your database. Paid plans start at $8/month.
What document parsing means here
“Parsing” is not one task. It usually means one of these:
- Classification: is this an invoice, a receipt, a contract or an ID?
- Field extraction: invoice number, date, vendor, line items, tax and total.
- Table extraction: rows and columns from a statement or report.
- Summarisation: parties, obligations, dates and termination terms from a contract.
- Validation: do the line items add up, and does the stated total match?
A single prompt can do several of these, but the output must be structured and checked. A plausible sentence is not a parsed record.
Model and limits
| Property | Value |
|---|---|
| Model ID | plugsky-parse |
| Context window | 262,144 tokens |
| Max output | 16,384 tokens |
| Input | Text, page images |
| Capabilities | Vision, JSON mode |
| Endpoint | OpenAI-compatible chat completions |
| Plans | Paid plans from Hobby $8/month |
A safe parsing pipeline
- Ingest: receive the PDF or image and store the original unchanged.
- Render: convert each PDF page to an image at a readable resolution, or extract text when the PDF already has a text layer.
- Parse: send one page at a time with the schema and the document type.
- Validate: required fields present, types correct, dates parseable, line items sum to the total.
- Escalate: route missing or contradictory fields to human review with the page image attached.
- Store: write the validated JSON plus the model name, prompt version and page hash for auditability.
Extract an invoice with JSON mode
import json
from openai import OpenAI
client = OpenAI(
api_key="YOUR_PLUGSKY_KEY",
base_url="https://api.plugsky.com/v1",
)
schema = {
"invoice_number": "string",
"invoice_date": "YYYY-MM-DD",
"vendor": "string",
"currency": "ISO 4217",
"line_items": [{"description": "string", "quantity": "number", "unit_price": "number", "amount": "number"}],
"subtotal": "number",
"tax": "number",
"total": "number",
}
response = client.chat.completions.create(
model="plugsky-parse",
response_format={"type": "json_object"},
messages=[
{"role": "system", "content": "Extract the invoice fields. Use null when a field is not present. Do not invent values."},
{"role": "user", "content": [
{"type": "text", "text": json.dumps({"schema": schema})},
{"type": "image_url", "image_url": {"url": "https://example.com/invoice-page-1.png"}},
]},
],
)
record = json.loads(response.choices[0].message.content)
assert record["total"] == round(record["subtotal"] + record["tax"], 2)
The assertion is the point. Do not trust the model's arithmetic; check it. If the total does not reconcile, send the page to review rather than correcting it silently.
What to measure before trusting the output
- Field-level accuracy on 100–300 real documents, not a vendor demo set.
- Layout coverage: scans, photos, rotated pages, multi-column tables and handwriting.
- Null behaviour: does the model return
nullfor a missing field, or invent a plausible value? - Arithmetic: line items vs subtotal vs tax vs total.
- Date and currency: ambiguous formats such as 03/04/2026 and currencies with different decimal conventions.
- Review rate: the share of documents a human must touch. That number drives the cost per document.
Cost per document
A one-page invoice is typically a small number of image tokens plus a JSON completion. At low volume, the plan price dominates; at high volume, the token count and the human-review rate dominate. Calculate the fully loaded cost:
cost_per_document =
(input_tokens × input_rate + output_tokens × output_rate)
+ (review_rate × minutes_per_review × reviewer_cost_per_minute)
For high-volume work, Builder ($60) and Scale ($120) have no monthly token cap, with fair-use concurrency and RPM. See pricing for the current plan caps.
Security and residency
Documents often contain personal or financial data. Decide before you send: which fields are required, whether the page can be redacted, how long the provider retains it, and where it is processed. Plugsky offers cloud, private VPC, on-prem and air-gapped deployment; confirm the contractual retention and residency terms for your plan before sending regulated documents.
FAQ
Can an LLM parse PDFs and invoices?
Yes, with a page image or extracted text and a strict schema. Validate required fields, totals and dates before writing to a system of record.
Which Plugsky model should I use for document parsing?
plugsky-parse, with a 262,144-token context and up to 16,384 output tokens. Text-only PDFs can also use a text model after extraction.
How accurate is invoice extraction?
It depends on scan quality, layout and handwriting. Measure field-level accuracy on your own set and route high-value fields to human review.
Does document parsing include data residency?
Plugsky offers cloud, private VPC, on-prem and air-gapped deployment. Residency and retention depend on the plan and contract; confirm before sending regulated documents.
Is document parsing included in the free plan?
The free plan includes Micro and Lite. plugsky-parse is available on paid plans from $8/month.
Reviewed 20 November 2026. Model limits are checked against the Plugsky model catalogue. Parsing accuracy varies by document; validate on your own corpus before production use.