Quick start¶

Steps 1 to 3 need only pip install contractex: no model, no network, no API key. Step 4 calls a model. It uses a local model through Ollama, which needs pip install "contractex[ollama]", Ollama running, and ollama pull llama3.1:8b.

The examples read examples/data/sample_nda.txt, a short fictitious NDA. Every block on this page is executed by the test suite, and the output shown is what it prints.

1. Parse the structure¶

Section numbering, defined terms and cross-references are found by deterministic parsing.

from contractex.structure import parse_structure

text = open("examples/data/sample_nda.txt", encoding="utf-8").read()
structure = parse_structure(text)

for section in structure.sections:
    print(f"{section.number} {section.title}".strip())
print(sorted(structure.defined_terms))
print("Unresolved:", [ref.raw_text for ref in structure.unresolved_refs])

MUTUAL NON-DISCLOSURE AGREEMENT
1 Definitions.
2 Obligations.
3 Term.
4 Exceptions.
5 Notices.
6 Governing Law.
['Agreement', 'Confidential Information', 'Harbourline', 'Kestrel', 'Purpose', 'Representatives']
Unresolved: ['Section 9']

The NDA refers to a Section 9 that does not exist. The parser reports it instead of guessing.

2. Find and redact personal data¶

from contractex.privacy import PIIDetector, PIIRedactor

detector = PIIDetector()
spans = detector.detect(text)
print("Presidio:", detector.using_presidio)
for span in spans:
    print(span.entity_type, span.text)

redacted = PIIRedactor().redact(text, spans)
print(next(line for line in redacted.text.splitlines() if line.startswith("Notices")))

Presidio: False
EMAIL_ADDRESS priya.raman@harbourline.test
PHONE_NUMBER 415-555-0132
EMAIL_ADDRESS legal@kestrelpoint.test
Notices to Harbourline go to Priya Raman at <EMAIL_ADDRESS_1> or <PHONE_NUMBER_1>. Notices to Kestrel go to <EMAIL_ADDRESS_2>.

This output comes from the regex fallback, which is what runs when Presidio is not installed. It catches the email addresses and the phone number but not the name "Priya Raman": the fallback has no name recogniser. Install contractex[privacy] if names must be redacted. See Limitations.

3. Chunk and trace a value to its source¶

ClauseAwareChunker cuts at section headings, and every chunk is an exact substring of the document. ProvenanceTracker resolves a value (for example, one an LLM returned) to character offsets in the original text.

from contractex.chunking import ClauseAwareChunker
from contractex.utils.provenance import ProvenanceTracker

chunks = ClauseAwareChunker(max_chunk_size=120, overlap=0).chunk(text)
for chunk in chunks:
    print(chunk.splitlines()[0])

tracker = ProvenanceTracker()
tracker.register_chunks(chunks, source_text=text)
span = tracker.find_span("governed by the laws of the State of Delaware")
print(span.char_start, span.char_end, text[span.char_start : span.char_end])

MUTUAL NON-DISCLOSURE AGREEMENT
1. Definitions.
3. Term.
6. Governing Law.
1245 1290 governed by the laws of the State of Delaware

4. Run tasks under a privacy profile¶

A LegalDoc carries a PrivacyProfile. Every built-in task that calls a model goes through the privacy router, which enforces the profile on each call.

from contractex import LegalDoc
from contractex.core.legal_document import DocType
from contractex.llm import LocalProvider
from contractex.privacy import PrivacyProfile
from contractex.tasks import TaskRegistry

doc = LegalDoc(
    doc_type=DocType.CONTRACT,
    full_text=text,
    privacy_profile=PrivacyProfile(sensitivity="restricted"),
)
llm = LocalProvider(model="llama3.1:8b")

pipeline = TaskRegistry.default().build_pipeline(
    ["pii_detection", "classification", "summarization"],
    task_kwargs={"summarization": {"llm_provider": llm}},
)
result = pipeline.run(doc)
print(sorted(result.extracted))
print(result.extracted["cuad_labels"])

['_task_timings', 'cuad_labels', 'pii_spans', 'summary']
['confidentiality', 'governing_law']

restricted means only a LocalProvider may see the document, and PII is redacted from each prompt before the call. The summary is model output and is not shown here.

If the same document were secret, no task that calls a model would run:

from contractex.privacy import PrivacyBlockedError

doc.privacy_profile = PrivacyProfile(sensitivity="secret")
try:
    pipeline.run(doc)
except PrivacyBlockedError as exc:
    print(type(exc).__name__)

PrivacyBlockedError

Next¶

Privacy model: the four sensitivity levels and what each enforces.
Tasks: which tasks call a model and what happens when the output is poor.
LLM providers: configuring OpenAI, Anthropic, Google or Ollama.