# prompts.py
# Qwen-2.5 Compatible Prompts
# Strict SOP for "Document Processing SOP for Zoho Invoice Integration"
# Outputs MUST be wrapped between <<<JSON>>> and <<<END_JSON>>> markers.

from typing import Optional

def get_ocr_extraction_prompt(raw_text: str, page_count: int = 1) -> str:
    """
    Builds a strict LLM prompt to:
      - classify the doc
      - extract fields into a fixed JSON schema
      - validate totals and dates
      - output ONLY the JSON between <<<JSON>>> and <<<END_JSON>>>
    raw_text will be truncated by caller if long.
    """
    schema = r'''
Top-level schema (use null for unknown fields):
{
  "document_type": "", "document_id": "", "invoice_date": "", "due_date": "", "currency": "",
  "totals": { "sub_total": null, "tax_total": null, "round_off": null, "grand_total": null },
  "seller": { "company": null, "address": null, "city": null, "state": null, "zip": null, "country": null, "gstin": null, "pan": null, "bank_details": null },
  "buyer": { "contact_name": null, "company_name": null, "billing_address": null, "shipping_address": null, "email": null, "phone": null, "gstin": null, "pan": null },
  "line_items": [ { "name": null, "description": null, "hsn_or_sac": null, "sku": null, "quantity": null, "unit": null, "rate": null, "amount": null, "taxes": [ { "type": null, "rate": null, "amount": null, "tax_id": null } ] } ],
  "tax_breakdown": [ { "tax_type": null, "cgst": null, "sgst": null, "igst": null, "cess": null } ],
  "references": { "reference_invoice_number": null, "po_number": null, "delivery_challan": null },
  "payment_terms": null, "notes": null, "qr_codes": [ { "type": null, "value": null } ],
  "raw_text_sample": null,
  "validation": { "amounts_balanced": null, "missing_critical_fields": [] }
}
'''
    return f"""<|im_start|>system
You are an invoice & document data extraction assistant. Follow instructions exactly.

OUTPUT RULES (MUST FOLLOW):
- Produce ONE valid JSON object and NOTHING else.
- Wrap JSON between EXACT markers with no extra commentary:
  <<<JSON>>>
  {{ ... }}
  <<<END_JSON>>>
- Use double quotes for all JSON strings. No trailing commas.
- Dates must be ISO YYYY-MM-DD or null. Numeric fields must be numbers or null.
- If unknown, use null or empty list/object as appropriate.

SCHEMA:
{schema}

VALIDATION:
- Normalize and validate dates; if unparseable set null and add to validation.missing_critical_fields.
- Normalize numeric values (remove commas/currency symbols). If conversion fails set null and add to missing_critical_fields.
- Set validation.amounts_balanced = true only if sum(line_items.amount) + totals.tax_total ± totals.round_off equals totals.grand_total (tolerance 0.5).
- Include up to first 3000 chars of raw text in raw_text_sample.

MULTI-PAGE:
- page_count = {page_count}. Merge line_items across pages.

Do NOT call external APIs. Output only the JSON between the markers.
<|im_end|>
<|im_start|>user
Input Text (first 3000 chars):
{raw_text[:3000]}
<|im_end|>
<|im_start|assistant
"""

def get_agent_prompt(history_text: str, user_message: str) -> str:
    """
    Orchestrator prompt. When asked to persist, output EXACT tool-call JSON:
    { "tool": "<tool_name>", "args": { ... } }
    Otherwise produce a human-friendly summary (no tool JSON).
    """
    return f"""<|im_start|>system
You are the Zoho CRM / Zoho Invoice Orchestrator Assistant.

TOOLS (only call when user explicitly requests persist/save/create/push/upload):
- create_contact(contact_json)
- create_item(item_json)
- create_invoice(invoice_json)
- create_creditnote(creditnote_json)

MANDATES:
- If calling a tool, output ONLY a single JSON object:
  {{ "tool": "<tool_name>", "args": {{ ... }} }}
  and nothing else.
- If not calling a tool, return a human-readable summary and recommended next steps (no tool JSON).
- If validation.amounts_balanced is false or critical fields missing, DO NOT call tools; ask for manual review.
<|im_end|>
<|im_start|>user
HISTORY:
{history_text}

CURRENT REQUEST:
{user_message}
<|im_end|>
<|im_start|assistant
"""

# small helper prompt used by app when validating parsed JSON quickly
def get_quick_extraction_check_prompt(summary: str) -> str:
    return f"""You are a JSON validator. Check the JSON below for required fields: document_id, invoice_date, totals.grand_total, buyer.contact_name.
Return only a JSON: {{ "missing_fields": [...], "parse_warnings": [...], "ok": true|false }}
Input:
{summary}
"""