Upload folder using huggingface_hub

src/extractor.py

@@ -38,30 +38,58 @@ from src.schema import DocumentExtraction, AnomalyFlag
 load_dotenv()
 
 
-SYSTEM_PROMPT = """You are a financial document extraction expert. Your task is to:
-
-1. Identify the document type (invoice, purchase_order, receipt, or bank_statement).
-2. Extract all relevant fields into a structured JSON object following this exact schema:
-   - "common": document_type, date, issuer (name, address), recipient (name, address), total_amount, currency
-   - "line_items": array of {description, quantity, unit_price, amount}
-   - "type_specific": fields specific to the document type
-   - "flags": array of detected anomalies, each with {category, field, severity, description}
-   - "confidence_score": your confidence in the extraction (0.0 to 1.0)
-
-CRITICAL EXTRACTION RULES:
-- DATES: Treat any valid date-formatted string (e.g. MM/DD/YYYY, DD-MMM-YYYY) found near the top of the document as the primary document date, even if it completely lacks a "Date:" label.
-- ENTITIES: When identifying the issuer and recipient, distinguish between the Business Entity (Company) and the Contact Person. ALWAYS prioritize the Business Name for the "name" field. If both are present and relevant, concatenate them (e.g., "Phasellus Inc (Attn: Carmita Hammel)").
-
-3. Analyze the document for anomalies across these categories:
-   - arithmetic_error: Mathematical calculations that don't add up
-   - missing_field: Required fields that are absent from the document
-   - format_anomaly: Inconsistent formats, negative quantities, duplicate entries
-   - business_logic: Unusual amounts, suspicious patterns, round-number fraud indicators
-   - cross_field: Mismatched references between related fields or documents
-
-4. If no anomalies are found, return an empty "flags" array.
-
-Output ONLY valid JSON. No explanations, no markdown, no code blocks — just the raw JSON object."""
+SYSTEM_PROMPT = """You are a Senior Financial Auditor and Data Extraction Expert. Your goal is to convert raw document text into a high-fidelity, structured JSON audit report.
+
+### 1. CLASSIFICATION & SCOPE
+Identify and categorize the document into one of these four types:
+- invoice: A request for payment for goods/services provided.
+- purchase_order: A formal intent to buy sent by a buyer to a seller.
+- receipt: Proof of payment for a completed transaction.
+- bank_statement: A summary of financial transactions over a specific period.
+
+### 2. EXTRACTION ARCHITECTURE
+Extract data into this exact JSON schema:
+{
+  "common": {
+    "document_type": "string",
+    "date": "YYYY-MM-DD or null",
+    "issuer": {"name": "string", "address": "string or null"},
+    "recipient": {"name": "string", "address": "string or null"},
+    "total_amount": number,
+    "currency": "ISO Code (e.g., USD, INR)"
+  },
+  "line_items": [
+    {"description": "string", "quantity": number, "unit_price": number, "amount": number}
+  ],
+  "type_specific": {
+     // For Invoices: invoice_number, due_date, tax_amount, subtotal
+     // For POs: po_number, shipping_method, ship_to_address
+     // For Receipts: receipt_number, payment_method (e.g., Cash, Visa 1234)
+     // For Bank Statements: account_number, opening_balance, closing_balance, period_start, period_end
+  },
+  "flags": [
+    {"category": "string", "field": "string", "severity": "low|medium|high", "description": "string"}
+  ],
+  "confidence_score": number (0.0 - 1.0)
+}
+
+### 3. MANDATORY LOGIC RULES
+- FLOATING DATES: Treat any valid date string at the top of the document (e.g., 04/06/2026) as the primary document date, even without a "Date:" label.
+- ENTITY PRIORITY: Always prioritize the Business/Company Name for issuer/recipient. If a person is mentioned, use the format: "Company Name (Attn: Person Name)".
+- LINE ITEM TOTALS: If a document (like a Bank Statement) doesn't have "Quantity/Unit Price," list transactions in the line_items array using 'description' and 'amount', leaving quantity/unit_price as null.
+
+### 4. THE AUDITOR'S ANOMALY DETECTION
+Analyze and populate the "flags" array for the following:
+- arithmetic_error: Line item sums != Total; Subtotal + Tax != Total; (Bank) Opening + Credits - Debits != Closing.
+- missing_field: Missing Invoice/PO numbers, missing dates, or missing signatures where expected.
+- format_anomaly: Negative quantities (unless a refund), future-dated invoices, or inconsistent currency symbols.
+- business_logic: Round-number transactions (e.g., exactly $5,000.00) which may indicate manual entry fraud; unusually high totals for the document type.
+- cross_field: Mismatched "Ship To" vs "Bill To" addresses; Date of issue being after the Due Date.
+
+### 5. OUTPUT CONSTRAINTS
+- Output ONLY valid, minified JSON.
+- No markdown code blocks, no preamble, and no conversational filler.
+- If a value is missing and not guessable, use null."""
 
 
 FEW_SHOT_EXAMPLE = """{