main.py

"""
FastAPI — Universal Order Confirmation PDF Extractor
Uses OpenAI GPT-4o to extract data from ANY supplier PDF format.

Run:
    pip install -r requirements.txt
    export OPENAI_API_KEY="sk-..."
    uvicorn main:app --reload --port 8000
"""

import io
import json
import os
import re
from typing import Any, List, Optional
from dotenv import load_dotenv
import pdfplumber
from openai import OpenAI, AuthenticationError as OpenAIAuthError
from fastapi import FastAPI, File, HTTPException, UploadFile
from pydantic import BaseModel, Field

load_dotenv()
# ─────────────────────────────────────────────────────────────────────────────
# Config
# ─────────────────────────────────────────────────────────────────────────────

OPENAI_API_KEY = os.environ.get("OPENAI_API_KEY", "")



# ─────────────────────────────────────────────────────────────────────────────
# Pydantic models
# ─────────────────────────────────────────────────────────────────────────────

class VoucherLine(BaseModel):
    type: str = Field(
        default="E3k.Web.Objects.DataTransfer.VoucherLines.ArticleVoucherLine, E3k.Web.Objects.DataTransfer",
        alias="$type",
    )
    Number: str                       # Supplier article number
    Quantity: float
    Price: float                      # Purchase unit price (from PDF)
    Description: str = ""
    VatCode: str = "01"
    DeliveryDate: Optional[str] = None

    model_config = {"populate_by_name": True}


class VoucherResponse(BaseModel):
    Supplier: str
    OurOrderNumber: str               # e.g. 2600385
    DeliveryDate: Optional[str]
    CustomerNumber: Optional[str]
    VoucherDate: Optional[str]
    Currency: str
    AdditionalFields: List[Any] = []
    VoucherLines: List[VoucherLine]


# ─────────────────────────────────────────────────────────────────────────────
# PDF text extraction
# ─────────────────────────────────────────────────────────────────────────────

def extract_text(pdf_bytes: bytes) -> str:
    parts = []
    with pdfplumber.open(io.BytesIO(pdf_bytes)) as pdf:
        for page in pdf.pages:
            t = page.extract_text()
            if t:
                parts.append(t)
    return "\n".join(parts)




# ─────────────────────────────────────────────────────────────────────────────
# LLM prompts
# ─────────────────────────────────────────────────────────────────────────────

SYSTEM_PROMPT = """You are a precise data extraction assistant for a Swiss hose service company (Schlauchservice Baumann GmbH).
You receive raw text extracted from supplier order confirmation PDFs (in German) and must extract structured order data.

IMPORTANT RULES:
- "OurOrderNumber" is the BUYER's order number — look for patterns like 2600xxx, BEST. 2600xxx, "Ihre Bestellung", "I/Bestellung", "Ihre Bestellnr.", "Ihr Auftrag", "Bestellreferenz"
- "CustomerNumber" is the supplier's customer number for Schlauchservice Baumann — look for "Kunden-Nr.", "Debitorennr.", "Kundennr.", "Kunden NR.", "Ihre Kunden-Nr.", "Kundennummer"
- All dates in the output must use the format DD.MM.YYYY (e.g. 18.03.2026). Convert ALL date formats to this.
- "DeliveryDate": the confirmed delivery/dispatch date — look for "Lieferung/Termin", "Auslieferdatum", "Versandtermin", "Lieferung", "Termin best.", "Warenausgangsdatum", "Versand-Datum", "Liefertermin"
  * If the delivery date is given as a calendar week like "KW 11" or "KW11", convert it to the WEDNESDAY of that ISO week in the document year. Example: "KW 11" in year 2026 → Wednesday of week 11, 2026 = 11.03.2026. Derive the year from the document/voucher date.
  * If no delivery date is mentioned at all, set DeliveryDate to null.
- "VoucherDate": the document/order confirmation date — look for "Datum", "Belegdatum", date next to "Auftragsbestätigung". Format as DD.MM.YYYY.
- For VoucherLines: extract ONLY real product/article lines. Skip shipping costs, surcharge lines, freight lines, and packaging lines UNLESS they have a real article number.
- "Number": the supplier's article/item number. If there are two numbers (our nr / their nr), use the SUPPLIER'S number (first one listed)
- "Price": unit purchase price. Use the net/discounted price if available. If only gross price with a discount %, calculate: price × (1 - discount/100)
- "Quantity": number of units ordered
- "Description": product description text
- "VatCode": always "01"
- "Currency": CHF or EUR
- If a delivery date is per-line, use the earliest confirmed date as the overall DeliveryDate too
- For delivery notes (Lieferschein): treat as order confirmation, extract what's available. Price may be 0 if not shown.

You must respond ONLY with a valid JSON object — no markdown, no explanation, no extra text."""

EXTRACTION_PROMPT = """Extract the order data from this supplier PDF text and return JSON matching this exact structure:

{{
  "Supplier": "supplier company name",
  "OurOrderNumber": "2600xxx",
  "DeliveryDate": "YYYY-MM-DD or null",
  "CustomerNumber": "customer number or null",
  "VoucherDate": "YYYY-MM-DD or null",
  "Currency": "CHF or EUR",
  "AdditionalFields": [],
  "VoucherLines": [
    {{
      "$type": "E3k.Web.Objects.DataTransfer.VoucherLines.ArticleVoucherLine, E3k.Web.Objects.DataTransfer",
      "Number": "article number",
      "Quantity": 1.0,
      "Price": 0.0,
      "Description": "product description",
      "VatCode": "01",
      "DeliveryDate": "DD.MM.YYYY or null"
    }}
  ]
}}

PDF TEXT:
---
{pdf_text}
---

Return ONLY the JSON object."""


# ─────────────────────────────────────────────────────────────────────────────
# LLM extraction via OpenAI API
# ─────────────────────────────────────────────────────────────────────────────

def llm_extract(pdf_text: str, api_key: str) -> dict:
    """Call OpenAI GPT-4o to extract structured order data from PDF text."""
    client = OpenAI(api_key=api_key)

    response = client.chat.completions.create(
        model="gpt-4o",
        max_tokens=4096,
        temperature=0,
        response_format={"type": "json_object"},   # enforces valid JSON output
        messages=[
            {"role": "system", "content": SYSTEM_PROMPT},
            {"role": "user",   "content": EXTRACTION_PROMPT.format(pdf_text=pdf_text)},
        ],
    )

    raw = response.choices[0].message.content.strip()

    # Strip any accidental markdown fences (safety net)
    raw = re.sub(r"^```(?:json)?\s*", "", raw)
    raw = re.sub(r"\s*```$", "", raw)

    return json.loads(raw)


def parse_pdf(pdf_bytes: bytes, api_key: str) -> VoucherResponse:
    # 1. Extract text from PDF
    pdf_text = extract_text(pdf_bytes)
    if not pdf_text.strip():
        raise ValueError("Could not extract any text from this PDF (may be image-based).")

    # 2. Send to LLM for extraction
    raw_data = llm_extract(pdf_text, api_key)

    # 3. Build VoucherLines
    voucher_lines = []
    for line in raw_data.get("VoucherLines", []):
        vl = VoucherLine(**{
            "$type": line.get(
                "$type",
                "E3k.Web.Objects.DataTransfer.VoucherLines.ArticleVoucherLine, E3k.Web.Objects.DataTransfer",
            ),
            "Number":       str(line.get("Number", "")),
            "Quantity":     float(line.get("Quantity", 0)),
            "Price":        float(line.get("Price", 0)),
            "Description":  line.get("Description", ""),
            "VatCode":      line.get("VatCode", "01"),
            "DeliveryDate": line.get("DeliveryDate"),
        })
        voucher_lines.append(vl)

    response = VoucherResponse(
        Supplier=raw_data.get("Supplier", "Unknown"),
        OurOrderNumber=str(raw_data.get("OurOrderNumber", "")),
        DeliveryDate=raw_data.get("DeliveryDate"),
        CustomerNumber=raw_data.get("CustomerNumber"),
        VoucherDate=raw_data.get("VoucherDate"),
        Currency=raw_data.get("Currency", "CHF"),
        AdditionalFields=[],
        VoucherLines=voucher_lines,
    )

    return response


# ─────────────────────────────────────────────────────────────────────────────
# FastAPI
# ─────────────────────────────────────────────────────────────────────────────

app = FastAPI(
    title="Universal Order Confirmation PDF Extractor",
    description=(
        "Upload any supplier order-confirmation PDF and receive ERP-ready JSON. "
        "Powered by GPT-4o — works with any supplier format."
    ),
    version="4.0.0",
)


@app.post(
    "/extract",
    response_model=VoucherResponse,
    summary="Extract order data from any supplier PDF",
)
async def extract_order(
    file: UploadFile = File(..., description="Supplier order-confirmation PDF")
):
    """
    Upload a PDF order confirmation from any supplier.
    GPT-4o extracts the order data and returns ERP-ready JSON.

    The `api_key` parameter defaults to the OPENAI_API_KEY environment variable.
    """
    api_key: str = OPENAI_API_KEY

    if not api_key:
        raise HTTPException(
            status_code=500,
            detail="OpenAI API key not configured. Set OPENAI_API_KEY env var or pass api_key param.",
        )

    if not file.filename.lower().endswith(".pdf"):
        raise HTTPException(status_code=400, detail="Only PDF files are accepted.")

    content = await file.read()
    try:
        result = parse_pdf(content, api_key)
    except json.JSONDecodeError as e:
        raise HTTPException(status_code=422, detail=f"LLM returned invalid JSON: {e}")
    except ValueError as e:
        raise HTTPException(status_code=422, detail=str(e))
    except OpenAIAuthError:
        raise HTTPException(status_code=401, detail="Invalid OpenAI API key.")
    except Exception as e:
        raise HTTPException(status_code=500, detail=f"Extraction error: {e}")

    return result


@app.get("/health", summary="Health check")
def health():
    return {"status": "ok", "version": "4.0.0"}