An LLM presents you a mistaken worth extracted from a doc as truth. That error is dear, particularly when basing choices on it. The Nutrient Data Extraction API targets that drawback. It parses PDFs, scans, photographs and Workplace recordsdata into spatial JSON or Markdown. Extract maps fields from a doc onto your JSON schema, and break up, classify and kind endpoints present helpful primitives on your doc pipelines. Every extracted worth is anchored to a web page through a bounding field, a match label and a confidence rating. Nutrient (previously PSPDFKit) runs the identical engine behind an MCP server, agent ability, Mac app and SDK.
TL;DR
- Dimension: A hosted API with 4 parse modes plus auto routing.
- Runs on: Nutrient cloud, your individual servers through SDK, or an Apple silicon Mac.
- The place to attempt it: Studio, a browser workspace to configure and check parsing, extraction,classification, splitting, and kind filling.
- Efficiency: 86.2 overall on Nutrient’s WholeDoc Bench, prime of 19 settings (reside API, Agentic).
- Greatest: 86.2 at 1.8¢ per web page, versus 67.1 at 17.8¢ for the perfect frontier-model setting.
How do you check it in Studio earlier than writing code?
Studio is Nutrient’s browser workspace for configuring and testing the API earlier than integrating it in your utility. It enables you to test parsing, extraction, classification, splitting, and kind filling by yourself recordsdata. A typical loop:
- Add a pattern doc and set the parameters you intend to ship.
- Draft a schema within the UI, or generate one based mostly by yourself paperwork.
- Run it on a contemporary doc, then examine values and citations.
- Save the setup as a versioned preset. The API calls it a processor, with a proc_ ID.
- Rating preset variations towards as much as 25 reviewed paperwork in an analysis set.
Processors saved in Studio can then be referred to as from the API by ID.
Which mode do you have to decide?
Auto routes every web page to construction, perceive or agentic. If the mode is omitted, the API defaults to understand.
| Operation | Greatest for | Value per web page | Extract per web page | WholeDoc Bench* |
| Parse: textual content | Born-digital PDFs and Workplace recordsdata, no OCR | $0.001 | $0.007** | Not examined |
| Parse: construction | OCR for scans with easy layouts, no AI | $0.0015 | $0.0075 | 39.4 |
| Parse: perceive | AI structure evaluation: types, advanced layouts | $0.009 | $0.015 | 80.8 |
| Parse: agentic | Imaginative and prescient language mannequin: cursive, degraded scans, charts | $0.018 | $0.024 | 86.2 |
| Classify | Zero-shot doc labels | $0.001 | n/a | n/a |
| Break up | Separating merged PDFs into paperwork | $0.003 | n/a | n/a |
*WholeDoc Bench parsing rating, 5 November 2026; not extraction accuracy. Extract provides $0.006 per web page. **API docs record text-mode extract; the agent ability begins at construction.
Pricing is in {dollars}, from 1 steadiness shared throughout Nutrient’s cloud APIs, so groups can combine and match. The free plan provides you $5 month-to-month utilization that covers about 330 understand-mode extractions. Starter prices $75 a month for $90 of utilization; Development prices $275 for $345.
How does grounded confidence work?
In Nutrient’s confidence post, a discipline scored 0.97 however got here from the mistaken column.
The present rating is a composite of non-obligatory alerts: token chance, candidate margin, grounding and format. Nutrient’s NLI grounding model is open beneath Apache-2.0. A grounded worth can nonetheless reply the mistaken discipline. When no rating exists, confidence is omitted, not low.
Every discipline additionally will get a match label: id_match, id_match_multiblock, id_match_partial, fuzzy_match or not_found. Gate on the label first, then confidence.
import json, requests
API_KEY = "your_api_key"
schema = {
"type": "object",
"properties": {
"invoice_number": {"type": "string", "description": "Invoice number below the supplier details."},
"total_amount": {"type": "number", "description": "Printed final total including tax. Not the subtotal."},
},
"required": ["invoice_number", "total_amount"],
}
directions = {"schema": schema, "parseConfig": {"mode": "understand"},
"options": {"includeCitations": True}}
with open("invoice.pdf", "rb") as f:
r = requests.put up(
"https://api.nutrient.io/extraction/extract",
headers={"Authorization": f"Bearer {API_KEY}"},
recordsdata={"file": ("invoice.pdf", f, "application/pdf"),
"instructions": (None, json.dumps(directions), "application/json")},
timeout=300,
)
r.raise_for_status()
out = r.json()["output"]
knowledge, metadata = out.get("data") or {}, out.get("metadata") or {}
EXACT = {"id_match", "id_match_multiblock"}
THRESHOLD = 0.97 # illustrative solely; consider on consultant paperwork
for discipline in schema["properties"]: # loop over anticipated fields so none skip evaluation
worth = knowledge.get(discipline)
meta = metadata.get(discipline) or {}
match, conf = meta.get("match"), meta.get("confidence") # None = no rating
passes = (worth will not be None and match in EXACT
and conf will not be None and conf >= THRESHOLD)
print(discipline, worth, match, conf, "page", meta.get("pageNumber"),
"PASSES_BASIC_CHECKS" if passes else "REVIEW")
This demo covers flat schemas solely; manufacturing wants sort and business-rule checks. The 0.97 bar mirrors cash fields in Nutrient’s mortgage tutorial.
How do classify and break up work?
A document decision model powers classify and split. Nutrient stories 96.2% break up F1 towards 63.7% for Typesafe AI’s Jev.
How do brokers use it?
The hosted MCP server wants no API key. Claude Code connects with 1 command:
claude mcp add --transport http --scope person nutrient https://mcp.nutrient.io/mcp
Claude and Codex additionally work through OAuth. Default agentic mode prices $0.018 per web page to parse.
Coding brokers can set up the document-extraction-api skill with npx abilities add pspdfkit-labs/nutrient-skills –ability document-extraction-api.
Can it run domestically?
On the Mac app, paperwork keep native solely with an area mannequin. Absolutely offline manufacturing wants the SDK.
How does it rating on WholeDoc Bench?
Nutrient’s new WholeDoc Bench scores entire paperwork, not single pages: 480 paperwork, 30+ languages, 7 areas.
| Setting | General | Tables | Charts | Place | ¢/web page |
| Nutrient Agentic | 86.2 | 87.1 | 64.5 | 88.6 | 1.8 |
| Nutrient Perceive | 80.8 | 86.1 | 35.1 | 86.4 | 0.9 |
| Claude Fable 5.1 (excessive effort) | 67.1 | 84.7 | 58.9 | 66.0 | 17.8 |
| Azure Doc Intelligence | 65.9 | 69.4 | 0 | 72.5 | ≈2.8 |
| GPT-6.1 Sol (medium effort) | 62.9 | 82.8 | 47.6 | 49.3 | ≈3.1 |
| Lengthen (efficiency) | 60.5 | 80.8 | 40.6 | 75.2 | 2.0 |
| Reducto (R-1) | 59.9 | 76.4 | 38.5 | 76.4 | ≈1.7 |
| LlamaParse Agentic Plus | 56.7 | 89.5 | 0* | 65.4 | ≈5.6 |
| AWS Textract | 49.9 | 71.4 | 0 | 70.0 | 1.5 |
| Docling | 40.0 | 63.9 | 20.2 | 0 | Native |
Chosen rows, WholeDoc leaderboard, 5 November 2026, record costs. *Rating nonetheless beneath evaluation.
Cross-page continuity is the largest hole. On 44 paperwork with page-spanning tables, 15 of 19 settings not often stored them entire.
Nutrient constructed and ran it, and the primary check set is non-public. The code is Apache-2.0, with a 10-document public pattern.
Key Takeaways
- Fields carry pages, areas, match labels and, when out there, confidence.
- Configure and check in Studio earlier than writing any API code.
- Perceive-mode extraction prices $15 per 1,000 pages.
- Agentic tops Nutrient’s WholeDoc Bench at 86.2 for 1.8¢ per web page.
- Charts stay onerous: 64.5 is the perfect WholeDoc chart rating.
See the Data Extraction API, benchmarks, MCP server and agent skill on GitHub.
Due to the Nutrient AI crew for the thought management / assets for this text. This text is sponsored.
The put up Nutrient Data Extraction API: JSON Matching Your Schema With Per-Field Citations and Confidence appeared first on MarkTechPost.

