Extract Obligations with Precision
ClauseMinds runs two extraction lanes. Deterministic rules handle dated obligations—renewal notices, termination deadlines, payment terms—where rules stay authoritative and cheap. A grounded LLM read covers what rules were never shaped to express: standing commitments like confidentiality, non-use, exclusivity, IP ownership, and insurance or transfer limits. Every LLM item quotes the exact source clause or is dropped, and nothing from that lane is accepted automatically.
The extraction pipeline
Upload a contract. ClauseMinds extracts text (with automatic OCR fallback for scanned pages), then runs both lanes: deterministic rules find dated obligations, and a grounded LLM read finds standing commitments plus a catch-all pass for anything else material. Both lanes produce candidates with a source snippet, page reference, and confidence band—and both land in the same review queue under the same trust standard.
- Multi-format support (PDF, DOCX, TXT, HTML, XLSX, images)
- Page-level source traceability for every extraction
- Confidence bands (high, medium_high, medium, low)
- Automatic OCR for scanned documents
- Structured normalized values for every extraction
- One review queue for both lanes—one trust standard
Renewal Notice
high“Either party may terminate upon 90 days written notice...”
Page 12Termination Period
medium high“Termination shall be effective 60 days after notice...”
Page 8Payment Term
high“Payment is due within Net 30 days of invoice date...”
Page 5Confidentiality (standing)
high“Recipient shall not disclose Confidential Information to any third party...”
Page 9
Built for accuracy and traceability
Two lanes, one standard: structured obligations you can check against the document—with full auditability from source clause to tracked deadline.
Source Traceability
Every rules-lane extraction links to the exact clause and page. Every LLM-lane item quotes the exact source clause—or it is dropped, not surfaced.
Confidence Scoring
Calibrated confidence bands (high, medium_high, medium, low) route low-confidence items to human review before tracking.
Multi-Format Support
PDF, DOCX, TXT, HTML, XLSX, and scanned images. Native text extraction where possible, OCR fallback when needed.
OCR Fallback
Automatic OCR for scanned pages with quality detection. Extractions from scans may have lower confidence—flagged for review.
Structured Output
Normalized value schemas per obligation type. Consistent fields for deadline computation, and governing-truth logic so amendments override the original.
Per-Type Profiles
Each document type declares what ClauseMinds tracks for it. An NDA is read for different commitments than a SaaS agreement—extraction follows the profile.
Two lanes, one review queue
Lane one is deterministic rules for dated obligations—the families where a date and a notice window are the whole story. Rules stay authoritative and cheap there. Lane two is a grounded LLM read for everything the rules were never shaped to express: standing commitments and a catch-all pass. Governing-truth logic handles amendments and addenda across both lanes.
- Dated obligations (rules): renewal notice windows, termination deadlines, payment terms
- Standing commitments (grounded LLM): confidentiality, non-use, exclusivity, IP ownership, insurance and transfer limits
- Catch-all read for material terms outside the named families
- Every LLM item quotes its source clause or is dropped; nothing from the LLM lane auto-accepts
- “Nothing found” is an explicit, reviewable claim—absence is recorded, not silent
- 📅
Renewal Notice — rules lane
90 days written notice before anniversary
- 📋
Termination Period — rules lane
60 days after written notice
- 💰
Payment Term — rules lane
Net 30 from invoice date
- 🔒
Confidentiality — LLM lane
Standing non-disclosure, quoted from clause 7.2
- 🤝
Exclusivity — LLM lane
Standing commitment, quoted from clause 4.1
See extraction in action
Upload a contract and get structured obligations from both lanes—dated deadlines and standing commitments—each with a source link, confidence score, and deadline tracking. Human review before anything is tracked.