83 lines
3.0 KiB
TypeScript
83 lines
3.0 KiB
TypeScript
|
import { NextRequest, NextResponse } from 'next/server';
|
||
|
|
import OpenAI from 'openai';
|
||
|
|
import { z } from 'zod';
|
||
|
|
import { zodResponseFormat } from 'openai/helpers/zod';
|
||
|
|
|
||
|
|
const openai = new OpenAI();
|
||
|
|
|
||
|
|
// Define the schema for individual items
|
||
|
|
const ItemSchema = z.object({
|
||
|
|
item: z.string().describe('name of item, transaction description, or entry'),
|
||
|
|
unit_price: z.number().describe('unit price, transaction amount, or individual cost'),
|
||
|
|
quantity: z.number().describe('quantity, count of items, or number of transactions (use 1 for single transactions)'),
|
||
|
|
sum: z.number().describe('total amount for this item, transaction total, or entry amount')
|
||
|
|
});
|
||
|
|
|
||
|
|
// Define the main extraction schema
|
||
|
|
const ExtractionSchema = z.object({
|
||
|
|
company: z.string().describe('name of company, bank, or organization'),
|
||
|
|
address: z.string().describe('address, location, or branch information'),
|
||
|
|
total_sum: z.number().describe('total amount, sum of transactions, total paid, or final balance'),
|
||
|
|
items: z.array(ItemSchema).describe('list of items, transactions, or entries')
|
||
|
|
});
|
||
|
|
|
||
|
|
export async function POST(request: NextRequest) {
|
||
|
|
try {
|
||
|
|
const { text, schema } = await request.json();
|
||
|
|
|
||
|
|
if (!text) {
|
||
|
|
return NextResponse.json(
|
||
|
|
{ error: 'No text provided' },
|
||
|
|
{ status: 400 }
|
||
|
|
);
|
||
|
|
}
|
||
|
|
|
||
|
|
const completion = await openai.beta.chat.completions.parse({
|
||
|
|
model: 'gpt-4o-2024-08-06',
|
||
|
|
messages: [
|
||
|
|
{
|
||
|
|
role: 'system',
|
||
|
|
content: `You are an expert at extracting structured data from documents.
|
||
|
|
You will be given text from a PDF document (likely an invoice or bank statement)
|
||
|
|
and should extract the required information into the given structure.
|
||
|
|
|
||
|
|
For bank statements:
|
||
|
|
- Treat transaction descriptions as items
|
||
|
|
- Use transaction amounts as unit prices
|
||
|
|
- Use 1 as quantity for single transactions
|
||
|
|
- Total amount can be total deposits, total checks, or final balance
|
||
|
|
|
||
|
|
For invoices:
|
||
|
|
- Extract traditional invoice fields
|
||
|
|
- Calculate missing values if possible (e.g., sum = unit_price * quantity)
|
||
|
|
|
||
|
|
For any document type:
|
||
|
|
- Look for organization names in headers or footers
|
||
|
|
- Find address or location information
|
||
|
|
- Identify itemized lists or transactions
|
||
|
|
- Look for total amounts or summaries
|
||
|
|
|
||
|
|
Be precise with numerical values, extract them as numbers, not strings.
|
||
|
|
If a value is missing but can be calculated, calculate it.
|
||
|
|
Ensure all required fields are filled with meaningful values.`
|
||
|
|
},
|
||
|
|
{
|
||
|
|
role: 'user',
|
||
|
|
content: text
|
||
|
|
}
|
||
|
|
],
|
||
|
|
response_format: zodResponseFormat(ExtractionSchema, 'document_extraction')
|
||
|
|
});
|
||
|
|
|
||
|
|
const extractedData = completion.choices[0].message.parsed;
|
||
|
|
|
||
|
|
return NextResponse.json(extractedData);
|
||
|
|
} catch (error) {
|
||
|
|
console.error('Error processing text:', error);
|
||
|
|
return NextResponse.json(
|
||
|
|
{ error: 'Error processing text' },
|
||
|
|
{ status: 500 }
|
||
|
|
);
|
||
|
|
}
|
||
|
|
}
|