Extract

One call#

curl https://api.wity.alphanimble.com/v1/extract \
  -H "Authorization: Bearer $WITY_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "text": "MUTUAL NON-DISCLOSURE AGREEMENT ... made as of the fourteenth day of March, 2025 ... between Kestrel Foods Ltd. ... and Brightmart Inc. ...",
    "fields": {
      "party": "a company or person that signs the agreement",
      "effective date": "the date the agreement takes effect",
      "term": "how long the confidentiality obligations last",
      "governing law": "the state or country whose law governs the agreement"
    },
    "types": { "effective date": "date", "term": "duration" }
  }'

{
  "fields": {
    "party": ["Kestrel Foods Ltd.", "Brightmart Inc."],
    "effective date": ["2025-03-14"],
    "term": ["three (3) years"],
    "governing law": ["New York"]
  },
  "entities": [
    { "field": "party", "text": "Kestrel Foods Ltd.", "start": 174, "end": 192, "confidence": 0.9929 },
    { "field": "party", "text": "Brightmart Inc.", "start": 267, "end": 282, "confidence": 0.9903 },
    { "field": "effective date", "text": "2025-03-14", "start": 134, "end": 135, "confidence": 0.9883 },
    { "field": "term", "text": "three (3) years", "start": 697, "end": 703, "confidence": 0.9985 },
    …
  ],
  "timing_ms": 407,
  "usage": { "input_tokens": 428, "output_tokens": 0 }
}

• fields lists each field's values in the order you asked for them. entities has the same values with their character offsets (text[start:end]) and a confidence from 0 to 1.
• Values are the document's own words: spacing is tidied, nothing is rephrased. A field with nothing in the document comes back as [].
• Typed fields (date, duration, number) come back normalised: "the fourteenth day of March, 2025" becomes 2025-03-14. Their start and end point at where in the document the value was read, as evidence.

Writing good fields#

• Say what the field is, not where it is. “the total amount to pay” works better than “the number at the bottom”.
• One kind of value per field. Ask for “a company that signs the agreement” and “the person who signs for a company” as two fields, not one.
• Name the role when the same kind of value plays several. On a receipt, “cash paid”, “change given back” and “tax amount” are all amounts: describe each one.
• Type anything you want normalised. Dates and terms were right much more often once typed.
• Ask for what you need. Each extra field adds a little time, and a field that has no answer in most documents tends to collect wrong ones.

How good it is#

Measured on the live API (20 October 2026) on public test sets we never tuned on, against hand-labelled values. A value counts as right when it shares at least half its words with the labelled one. Times are for one request at a time; requests running together share the server.

Where it is weaker

• Values that need reasoning rather than reading, unless the field is typed (a term written as an end date, for instance).
• Two roles of the same kind in one document: signatories taken for parties, or cash for change.
• Very long lists: at most 30 values per field.
• Conventions: whether a value includes its clause, or “Consac LLC” versus “Consac, LLC”.
• The confidence is a good ranking of values, but it has not been calibrated as a probability.

Good to know#

• Limits: up to 60,000 characters and 16 fields per request; field names up to 64 characters, descriptions up to 300.
• Billing: usage.input_tokens is the document and the field list, counted once, at the usual $0.042 per million. Requests that fail cost nothing.
• Extract after split: cut a bundle into items with split, then extract the same fields from every item.
• Full reference: POST /v1/extract.