Stage 3 — OCR & Data Extraction

Goal

Convert raw evidence collected in Stage 2 into structured, machine-readable document data. This stage focuses only on extracting information from documents—it does not perform legal interpretation.

Output: Document Intelligence Package.

Scope

This stage is responsible for image preprocessing, OCR, layout extraction, field extraction, confidence scoring, document-level metadata and summaries.

This stage does not classify documents, build chain of title, detect liens or perform legal reasoning.

Processing Pipeline

Raw Document
      │
      ▼
Preprocessing
      │
      ▼
Tesseract OCR
      │
      ▼
Layout Extraction
      │
      ▼
LangChain Extraction Pipeline
      │
      ▼
Schema Validation
      │
      ▼
Confidence Scoring
      │
      ▼
Document Summary
      │
      ▼
Structured Document

OCR

Initial implementation:

• Tesseract OCR

Preprocessing:

• Deskew
• Rotate
• Denoise
• Contrast enhancement
• TIFF conversion
• Blank page detection

Intelligent Extraction

Field extraction should not rely primarily on regex or fixed templates.

Instead, use a LangChain-based extraction pipeline that converts OCR output into structured JSON.

Pipeline

• OCR Text
• Page Layout
• Bounding Boxes
• Document Metadata ↓ LangChain ↓ Prompted LLM ↓ Structured JSON ↓ Schema Validation ↓ Confidence Scoring

The LLM should extract all supported fields into a deterministic schema.

Validation should verify:

• Required fields
• Data types
• Field formats
• Cross-field consistency

Regex should only be used as a validator or fallback for deterministic values such as:

• Dates
• Instrument Numbers
• APNs
• Recording Numbers
• Book/Page references

It should never be the primary extraction mechanism.

Field Extraction

Every document should expose a structured schema instead of only OCR text.

Common Fields

• Document ID
• Document Type (unknown until Stage 4)
• Recording Date
• Execution Date
• Recording Office
• Instrument Number
• Book
• Page
• Document Number
• Page Count
• County
• State
• Confidence Score

Property Fields

• Property Address
• Parcel / APN
• Legal Description
• Lot
• Block
• Subdivision

Parties

• Grantor(s)
• Grantee(s)
• Buyer(s)
• Seller(s)
• Borrower(s)
• Lender(s)
• Trustee
• Beneficiary
• Notary
• Witnesses

Financial Fields

• Loan Amount
• Consideration Amount
• Taxes
• Fees

References

• Previous Instrument
• Related Instrument
• Referenced Mortgage
• Referenced Deed

Signatures

• Signature Present
• Notary Present
• Seal Present

Layout Information

Store:

• OCR text
• Bounding boxes
• Lines
• Paragraphs
• Pages
• Confidence by region

Document Summary

Generate a concise summary for every document.

Example:

• Warranty Deed transferring ownership from John Doe to Jane Doe.
• Recorded on 2022-01-15.
• References Parcel XXXXX.

This summary is attached to the document only.

Cumulative Summary

Do not build a property-level cumulative summary in Stage 3.

Stage 3 should only generate document-level summaries.

A cumulative/property-level summary should be generated later, after:

• Document Classification
• Chain of Title
• Lien Search
• Risk Analysis

At that point the system has enough context to produce an accurate property summary.

Structured Document Output

Each processed document should contain:

• Raw File
• Preprocessed Images
• OCR Text
• Layout Information
• Bounding Boxes
• Structured JSON
• Extracted Fields
• Validation Results
• Confidence Scores
• Document Summary
• Processing Logs
• Extraction Version
• Prompt Version

Edge Cases

• Handwritten notes
• Poor scans
• Rotated pages
• Stamps
• Multi-column layouts
• Mixed document types
• Missing pages
• Corrupted PDFs

Human Involvement

None.

Low-confidence extraction is preserved and forwarded.

Definition of Done

• OCR completed.
• Structured fields extracted.
• Layout preserved.
• Document summary generated.
• Confidence scores generated.
• Structured document created.
• Ready for Stage 4.