Parts and Sublet List Extraction

Automated extraction of repair assessment data from Parts and Sublet List PDFs using Claude AI Vision.

System: Invoice Extraction Server (Node.js + Claude API) Integration: Zoho Creator → Extraction Server → Zoho Creator Processing time: ~30-90 seconds per PDF (depends on page count)

Overview

The Parts Assessment extraction workflow processes Parts and Sublet List documents generated by assessment/quoting systems. These documents detail parts needed for vehicle repairs, including labor, sublet work, and parts breakdowns.

What It Extracts

Header Fields:

Field	Description	Example
`assessor_name`	Name of assessor/quoter	"John Smith"
`assessment_date`	Date of assessment	"15/07/2026"
`claim_number`	Insurance claim reference	"CLM-2024-001"
`vehicle_make`	Vehicle manufacturer	"Toyota"
`vehicle_model`	Vehicle model	"Camry"
`vehicle_year`	Model year	"2020"
`vin`	Vehicle Identification Number	"JTDKARFU5L3123456"
`registration`	Vehicle registration	"ABC123"
`customer_name`	Customer/claimant name	"Jane Doe"

Breakdown Sections:

The document typically contains three main sections:

Parts List - Repair parts needed
Labour - Labor hours and rates
Sublet Work - External services (painting, glass, etc.)

Line Item Structure:

Field	Description	Example
`description`	Item description	"Front Bumper Cover"
`part_number`	Manufacturer part number	"52119-06F00"
`quantity`	Quantity required	1
`unit_price`	Unit price ex GST	450.00
`total`	Line total ex GST	450.00
`category`	Parts, Labour, or Sublet	"Parts"

Summary Totals:

Field	Description
`parts_total`	Total parts cost ex GST
`labour_total`	Total labour cost ex GST
`sublet_total`	Total sublet cost ex GST
`subtotal`	Combined subtotal ex GST
`gst`	GST amount (10%)
`total`	Grand total inc GST

Process Flow

graph TD
    A[User uploads Assessment PDF] --> B[Zoho Creator webhook triggers]
    B --> C[Server: Download PDF]
    C --> D[Convert PDF to images]
    D --> E[Extract with Claude Vision]
    E --> F[Parse and validate JSON]
    F --> G[Normalize nested structures]
    G --> H[Send to Zoho API]
    H --> I[Zoho creates/updates record]
    I --> J[Complete]

Step-by-Step Process

1. File Upload & Trigger

Who: Assessors, Quoters, Parts team Where: Zoho Creator → Parts Assessment upload form What happens:

User uploads Parts and Sublet List PDF
Zoho Creator webhook fires with:
- Record ID (upload_id)
- File URL
- Environment (dev/production)

2. PDF Download

System: Invoice Extraction Server What happens:

Downloads PDF from Zoho Creator using OAuth token
Saves to temporary directory: temp/partsassessment-{timestamp}/
Retry mechanism: 3 attempts with 3-second intervals

:::info Download Retry Only File downloads retry automatically (3 attempts, 3s delay). Claude AI and Zoho API calls do NOT retry - if they fail, manual intervention is required. :::

3. PDF to Image Conversion

Technology: pdf2pic + GraphicsMagick What happens:

PDF split into individual page images (PNG format)
High-resolution conversion for optimal Vision AI accuracy
All pages converted (multi-page documents supported)
Output: Array of page images in temp/{job_id}/pages/

Why images instead of direct PDF? Parts and Sublet Lists often have complex formatting, tables, and visual layouts. Image-based extraction provides better accuracy for these structured documents.

4. Data Extraction (Claude Vision)

Model: Claude Opus 4 (Anthropic) Input: Array of page images (base64-encoded) What happens:

All page images sent in single API call
Claude Vision analyzes entire document layout
Extracts header, parts list, labour, sublet sections
Returns structured JSON

Prompt engineering:

Australian assessment format handling
Part number extraction from descriptions
Labour rate and hours parsing
Sublet work identification
Multi-page document handling

5. Response Normalization

Challenge: Claude may return nested structures

Sometimes Claude returns nested objects like:

{
  "header": {
    "assessor_name": "John Smith",
    "claim_number": "CLM-001"
  },
  "vehicle": {
    "make": "Toyota",
    "model": "Camry"
  },
  "parts": [...]
}

Solution: Automatic flattening

// Flatten nested header
if (extractedData.header && typeof extractedData.header === 'object') {
  const { header, ...rest } = extractedData;
  extractedData = { ...header, ...rest };
}

// Flatten nested vehicle
if (extractedData.vehicle && typeof extractedData.vehicle === 'object') {
  const { vehicle, ...rest } = extractedData;
  extractedData = { ...vehicle, ...rest };
}

Result: Flat structure expected by Zoho API

6. JSON Parsing & Validation

What happens:

Claude response may be wrapped in markdown code blocks:
```
```json
{ ... }
```
```
Server strips markdown wrappers
Parses JSON
Validates required fields
Checks data types

7. Send to Zoho API

Endpoint: Custom Zoho Creator API Environment-aware:

Development: poc_extract_parts_details
Production: po_pdf_extract_details_prod

Payload:

{
  "upload_id": "1234567890",
  "status": "success",
  "extracted_data": "{...JSON string of all fields...}"
}

No retry mechanism - single attempt

8. Response & Cleanup

What happens:

Zoho processes extracted data
Updates assessment upload record
Server cleans up temporary files and images
Returns success/failure status

Extraction Prompt Details

The Parts Assessment prompt instructs Claude to:

Document Structure Recognition

Identify header section (top of first page)
Locate Parts list table
Find Labour section
Identify Sublet work section
Extract summary totals (bottom of document)

Part Number Handling

Similar to PO PDF extraction, part numbers may be embedded in descriptions:

"Front Bumper Cover 52119-06F00"
 └─ Description ─┘ └─ Part Number ─┘

Extraction should split into:

description: "Front Bumper Cover"
part_number: "52119-06F00"

Date Formatting

Always convert to dd/MM/yyyy format
Handle various Australian date formats

Numeric Values

Remove currency symbols ($, AUD)
Parse as decimal numbers
GST = 10% of subtotal
Validate totals match sum of line items

Multi-Page Handling

Extract all pages
Continuation tables across pages
Maintain correct item sequence

Error Handling

Error Type	Handling
File download failure	Retry 3 times with 3s delay, then fail
PDF conversion error	Fail immediately with error message
Claude API error	Fail immediately, log error, no retry
JSON parsing error	Fail with parse error details
Zoho API error	Log error, no retry
Nested structure	Auto-flatten, log normalization

Environment Configuration

Environment	Zoho API	API Key
Development	`poc_extract_parts_details`	`DK6ThON2OyGvTxFkg24phyQ8a`
Production	`po_pdf_extract_details_prod`	`1FCKNxWAhzN7yPB7gUnYZTe9O`

:::info Shared API Endpoints Parts Assessment and PO PDF extraction use the same Zoho API endpoints. The system differentiates between them based on the upload form used. :::

Common Data Patterns

Labour Line Items

Labour is typically broken down by operation:

Description	Hours	Rate	Total
Remove & Refit Front Bumper	1.5	85.00	127.50
Paint Front Bumper	2.0	85.00	170.00

Sublet Work

External services billed separately:

Description	Quantity	Unit Price	Total
Windscreen Replacement	1	450.00	450.00
Headlight Restoration	2	120.00	240.00

Parts List

Standard parts breakdown:

Part	Part Number	Qty	Unit Price	Total
Front Bumper Cover	52119-06F00	1	450.00	450.00
Fog Light LH	84501-06140	1	180.00	180.00

Data Quality Checks

Automatic Validations

Total reconciliation
- Parts total = sum of parts lines
- Labour total = sum of labour lines
- Sublet total = sum of sublet lines
- Subtotal = parts + labour + sublet
- GST = subtotal × 0.10
- Total = subtotal + GST
Required fields
- Assessor name
- Vehicle details
- At least one line item
Field formats
- Dates: dd/MM/yyyy
- VIN: 17 characters (if present)
- Numbers: valid decimals

Technical Details

Processing method: Image-based Vision AI extraction Model: Claude Opus 4 (Anthropic) Max tokens: 16,000 Image format: PNG (high resolution) Timeout: 30 seconds per document

Dependencies:

@anthropic-ai/sdk - Claude AI Vision API
pdf2pic - PDF to image conversion
sharp - Image processing
fs - File system operations
axios - HTTP client for Zoho API

Best Practices

For Assessment Team

Export quality matters - Clear, high-resolution PDFs extract better
Standard formats work best - Common assessment templates have higher accuracy
Multi-page OK - System handles documents with many pages
Review extractions - Verify key totals and part numbers in Zoho

For IT/Developers

Monitor image conversion - Check GraphicsMagick logs for errors
Watch for nested responses - Flattening logic catches these automatically
Track API usage - Claude Vision tokens consumed per page
Test both environments - Dev and prod use same endpoints but different keys

Troubleshooting

Issue	Cause	Solution
PDF conversion fails	GraphicsMagick not installed	Verify Docker container has `graphicsmagick` package
Nested structure in response	Claude's interpretation	Auto-flattening handles this (already implemented)
Missing line items	Poor PDF quality, complex layout	Request clearer export, higher resolution
Incorrect totals	Calculation mismatch in source	Manual review and correction
Part numbers in description	No splitting in source	Prompt instructs separation, may need manual cleanup

Performance Metrics

Average processing times:

1-page document: ~30 seconds
3-page document: ~60 seconds
5+ page document: ~90 seconds

Accuracy expectations:

Header extraction: 95%+
Line item extraction: 90%+
Total calculation: 98%+

Invoice Extraction - Bulk invoice processing
PO PDF Extraction - Cost Price Report extraction
Reference: Order Parts Form - Parts data structure

For technical support or issues, contact the development team.

Overview​

What It Extracts​

Process Flow​

Step-by-Step Process​

1. File Upload & Trigger​

2. PDF Download​

3. PDF to Image Conversion​

4. Data Extraction (Claude Vision)​

5. Response Normalization​

6. JSON Parsing & Validation​

7. Send to Zoho API​

8. Response & Cleanup​

Extraction Prompt Details​

Document Structure Recognition​

Part Number Handling​

Date Formatting​

Numeric Values​

Multi-Page Handling​

Error Handling​

Environment Configuration​

Common Data Patterns​

Labour Line Items​

Sublet Work​

Parts List​

Data Quality Checks​

Automatic Validations​

Technical Details​

Best Practices​

For Assessment Team​

For IT/Developers​

Troubleshooting​

Performance Metrics​

Related Documentation​