Skip to content

Contract & document extraction

Goal: turn a folder of PDF contracts into a spreadsheet of structured fields.

The pipeline

smartpipe map "Extract {party_a, party_b, effective_date, total_value number, governing_law}" 'contracts/*.pdf' --output csv > contracts.csv

Each PDF becomes one item; smartpipe parses it to text automatically, extracts the five fields, and writes a CSV you can open in Excel.

party_a,party_b,effective_date,total_value,governing_law
Acme Corp,Globex LLC,2025-03-01,250000,Delaware
Initech,Umbrella Inc,2025-06-15,89000,California

Enforce the shape for production

Inline braces are convenient for exploration. For a pipeline you run every week, pin a JSON Schema so types are enforced and stray fields are dropped:

// contract.json
{
  "type": "object",
  "properties": {
    "party_a": { "type": "string" },
    "party_b": { "type": "string" },
    "effective_date": { "type": "string" },
    "total_value": { "type": "number" },
    "governing_law": { "type": "string" }
  },
  "required": ["party_a", "party_b"],
  "additionalProperties": false
}

smartpipe map "Extract the contract details" 'contracts/*.pdf' --schema contract.json --output csv > contracts.csv

Now total_value is a real number, and a document that doesn't parse cleanly is retried once, then skipped with a warning - so a single bad file doesn't stop the batch.

Narrow to the contracts you care about first

Chain a semantic filter to process only the relevant documents:

smartpipe filter "is a signed vendor agreement" 'docs/**/*.pdf' \
| smartpipe map "Extract {vendor, renewal_date, annual_cost number}" --from-files --output csv

The filter emits the paths of matching files; --from-files feeds those paths to map. You extract fields from only the vendor agreements, skipping everything else.

Save it as a stage

Once the extraction works, freeze it as an executable .sem file so the whole team runs the same stage:

#!/usr/bin/env -S smartpipe run
verb = "map"
prompt = "Extract {vendor, renewal_date, annual_cost number}"
schema-file = "contract.json"
output = "csv"

chmod +x extract-contract.sem
smartpipe filter "is a signed vendor agreement" 'docs/**/*.pdf' \
| ./extract-contract.sem --from-files > contracts.csv

schema-file resolves next to the .sem file, so the script and its schema travel together. Full format: .sem files.

See also