Comparative Estimation Study

Compare AI-generated baseline estimations with actual company PDFs to identify pricing differences and detect new items.

When to Use This Skill

Use this skill when the user asks about:

• Comparing AI estimations with company quotes
• Validating baseline estimations against actual PDFs
• Finding price differences between baseline and company data
• Identifying new items in company estimates
• Reconciling estimation data with supplier quotes

Function Signature

def process_comparative_study(file_urls: List[str], output_path: str) -> str:
    """
    Compare baseline JSON estimation with company PDF quotes

    Matching Rules:
    - Two-layer designation matching: string comparison → semantic similarity (embeddings)
    - Pre-filters: number + unit + quantity must match
    - is_new = False: For all items that exist in baseline JSON
    - is_new = True: ONLY for items in PDF with 'number' not in baseline
    - Prices updated if different or if baseline has null values
    - Section totals automatically recalculated

    Args:
        file_urls: List of URLs (first must be JSON baseline, rest are PDFs)
        output_path: Path to save output JSON

    Returns:
        str: Path to output JSON file with is_new flags (includes UUID in filename)

    Raises:
        ValueError: If no JSON or no PDFs found in URLs
        Exception: For download or processing errors
    """

CLI Usage

# Compare baseline JSON with company PDFs
python comparative_study.py \
  --file-urls https://storage.com/baseline.json https://storage.com/quote1.pdf https://storage.com/quote2.pdf \
  --output-path temp_files/temp_project_123/comparison_result.json

# Note: The script automatically appends a UUID to the output filename
# Example output: comparison_result_a1b2c3d4-5678-90ab-cdef-1234567890ab.json

⏱️ Important: When using this skill programmatically (not CLI), ensure timeout is set to at least 3 minutes to allow sufficient time for PDF extraction and semantic matching operations.

Input Format

File URLs (Required Order)

1. First URL: Baseline JSON (AI estimation output)
2. Remaining URLs: Company PDF quotes (1 or more)

Baseline JSON Structure

{
  "projectId": "123",
  "projectName": "Construction Project",
  "pricingLines": [
    {
      "type": "line",
      "number": "1.1",
      "designation": "Installation de chantier",
      "unit": "F",
      "quantity": "1",
      "unitPrice": 24000,
      "totalPrice": 24000
    }
  ]
}

PDF Content

Company estimate documents with pricing tables containing:

• Designation (item description)
• Unit (measurement unit)
• Quantity
• Unit price
• Total price

Matching Logic

Two-Layer Designation Matching

Pre-filters (must pass first):

1. Number: Must match or either is null (e.g., "1.1" = "1.1")
2. Unit: Must match or either is null (e.g., "F" = "F")
3. Quantity: Must match or either is null (e.g., "1" = "1")

Layer 1 - String Matching (Fast):

• Designation matches via string comparison (exact or substring)
• Normalized text (lowercase, no accents, no special chars)

Layer 2 - Semantic Matching (Fallback):

• If string match fails, use Voyage AI embeddings
• Computes cosine similarity between designations
• Threshold: 0.8 (80% semantic similarity)
• Example: "Installation de chantier" ≈ "Installation chantier principal"

Price Update Logic

• Baseline has null, PDF has values: Always update with PDF values
• Both have values and differ: Update with PDF values
• Both have same values: Keep baseline values unchanged

is_new Flag Rules

• is_new = False: All items that exist in baseline JSON (by number)
• is_new = True: ONLY items in PDF whose 'number' doesn't exist in baseline
• Applies to all types: line, section, titre

Special Features

• Null value handling: Null/empty values treated as wildcards in matching
• Section totals: Automatically recalculated for type="titre" items
• Type flexibility: Matches across different types (section → line)

Output Format

Updated JSON with is_new field added to each pricing line:

{
  "projectId": "123",
  "pricingLines": [
    {
      "type": "line",
      "number": "1.1",
      "designation": "Installation de chantier",
      "unit": "F",
      "quantity": "1",
      "unitPrice": 24000,
      "totalPrice": 24000,
      "is_new": false
    },
    {
      "type": "line",
      "number": "1.2",
      "designation": "Signalisation de chantier",
      "unit": "F",
      "quantity": "1",
      "unitPrice": 5000,
      "totalPrice": 5000,
      "is_new": true
    }
  ]
}

Processing Pipeline

1. Download Files - Fetches JSON and PDFs from URLs
2. Load Baseline - Parses baseline JSON estimation
3. Extract PDFs - Uses LlamaExtract to parse PDF pricing tables
4. Merge PDF Data - Combines pricing lines from all PDFs
5. Build Baseline Numbers - Collects all 'number' fields from baseline (all types)
6. Match & Update - Applies two-layer matching (string → embeddings) and updates prices
7. Add New Items - Adds PDF items whose numbers don't exist in baseline
8. Recalculate Totals - Updates section/titre totalPrice based on child lines
9. Save Output - Writes updated JSON with is_new flags (UUID automatically appended to filename)

Error Handling

• No JSON URL ValueError
• No PDF URLs ValueError
• Download failure Exception with URL details
• Invalid file format Logs error and skips
• PDF extraction failure Exception

Environment Variables

# Required for PDF extraction
LLAMA_CLOUD_API_KEY=llx-...   # For LlamaExtract PDF parsing

# Required for semantic similarity matching
VOYAGE_API_KEY=pa-...          # For Voyage AI embeddings

Examples

Programmatic Usage

from comparative_study import process_comparative_study

file_urls = [
    "https://storage.com/baseline.json",
    "https://storage.com/company_quote1.pdf",
    "https://storage.com/company_quote2.pdf"
]
output_path = "temp_files/temp_project_456/comparison.json"

result = process_comparative_study(file_urls, output_path)
# Returns: "temp_files/temp_project_456/comparison_<uuid>.json"
# Example: "temp_files/temp_project_456/comparison_a1b2c3d4-5678-90ab-cdef-1234567890ab.json"

Notes

• URL Order: JSON must be first, followed by PDFs
• Multiple PDFs: All PDF pricing lines are merged before matching
• Text Normalization: Removes accents, special chars, common French stop words
• Price Tolerance: 1% tolerance for floating-point precision
• Temporary Files: Auto-cleaned after processing
• Designation Priority: Uses PDF designation when is_new = True