Skip to content

Contract analysis

Three steps, only one of which calls a model:

  1. Structure (no model): sections, defined terms, cross-references, recitals and a governing-law hint, from deterministic parsing.
  2. Extraction (model): ContractExtractor produces a typed Contract with parties, clauses, financial terms and risks.
  3. Playbook review (no model): keyword rules and required-clause checks over the extracted clauses.

Structure

from textwrap import shorten

from contractex.structure import parse_structure

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

term = structure.defined_terms["Confidential Information"]
print(term.definition_section, "|", shorten(term.definition_text, 60))
print([(ref.raw_text, ref.resolved) for ref in structure.cross_references])
print(structure.governing_law_hint, "|", shorten(structure.recitals[0], 50))

1 | any non-public information disclosed by one party to [...]
[('Section 4', True), ('Section 2', True), ('Section 9', False)]
Delaware | the parties wish to explore a possible data- [...]

governing_law_hint is a regular-expression signal. It misses phrasings such as "Delaware law governs this Agreement"; the evaluation guide measures exactly that.

Extraction

ContractExtractor loads a document (PDF, DOCX or text) or takes text directly, chunks it with ClauseAwareChunker, and makes schema-constrained model calls. Document details come from the first two chunks, capped at 12,000 characters. Clauses and financial terms come from one call per chunk. Risks come from keyword rules plus one call over the first 20,000 characters. It needs a provider, and it has no document privacy profile. Pass a guarded provider, or use the contract_extraction task, which does that for you.

from contractex import ContractExtractor
from contractex.llm import LocalProvider

extractor = ContractExtractor(llm_provider=LocalProvider(model="llama3.1:8b"), confidence_threshold=0.7)
contract = extractor.extract("examples/data/sample_nda.txt")
print(type(contract).__name__, contract.metadata.filename)

Contract sample_nda.txt

What the model returns is not shown here. Items below confidence_threshold are kept and listed in contract.metadata.warnings, never silently dropped.

Playbook review

A playbook lists required clause types and rules. RiskAnalyzer runs it over extracted clauses; here two clauses are built by hand, to show the deterministic part on its own:

from contractex.analysis import RiskAnalyzer
from contractex.core.models import Clause, Contract
from contractex.playbooks import StandardNDAPlaybook

extracted = Contract(
    clauses=[
        Clause(
            clause_type="confidentiality",
            section_number="2",
            text="Each party shall keep the other party's Confidential Information secret "
            "and use it only for the Purpose.",
        ),
        Clause(
            clause_type="governing_law",
            section_number="6",
            text="This Agreement is governed by the laws of the State of Delaware.",
        ),
    ]
)

analyzer = RiskAnalyzer(playbook=StandardNDAPlaybook())
for flag in analyzer.analyze(extracted):
    print(flag.severity.value, flag.playbook_rule, "section", flag.clause_ref)
print("Missing:", analyzer.missing_clauses(extracted))

medium nda.scope.overly_broad section 2
medium nda.return_of_info.missing section 2
Missing: ['termination_for_cause', 'effective_date', 'expiration_date']

Rules are keyword checks on each clause. The overly_broad flag fires because clause 2 contains no carve-out wording, even though this NDA puts its carve-outs in Section 4. Treat flags as prompts for review, not findings. SaaSPlaybook is also built in. Playbooks round-trip to YAML with Playbook.to_yaml() and Playbook.from_yaml().

Export

Contract is a Pydantic model: contract.model_dump_json() gives JSON. contract.to_dataframe() and contract.to_excel(path) need contractex[export].

print(extracted.model_dump_json(include={"clauses": {0: {"clause_type", "section_number"}}}))

{"clauses":[{"clause_type":"confidentiality","section_number":"2"}]}