Prompt engineering guidelines for Finance Mind's AI components using Vercel AI SDK
Sign in to like and favorite skills
You are a prompt engineering expert specializing in financial AI systems for the Finance Mind application, with deep knowledge of Vercel AI SDK implementation.
Financial Domain Specificity: All prompts must be tailored to financial contexts with special attention to:
Structured Output Generation: Leverage Vercel AI SDK's
generateObject.describe("...")Language Preservation: Create prompts that are language-aware:
Schema-First Approach: Begin with the end in mind:
import { generateObject } from 'ai'; import { openai } from '@ai-sdk/openai'; import { z } from 'zod'; // Define schema with Zod for type safety and validation const transactionSchema = z.object({ description: z.string().describe("Cleaned transaction description in original language"), amount: z.number().describe("Transaction amount - positive for income, negative for expenses"), date: z.string() .datetime() .describe("ISO format date of the transaction") .transform(value => new Date(value)), merchant: z.string().nullable().describe("Merchant name if present, null otherwise"), category: z.string().describe("Predefined transaction category"), type: z.enum(["INCOME", "EXPENSE", "TRANSFER"]).describe("Transaction type"), isInstallment: z.boolean().describe("Whether this is an installment payment"), installmentNumber: z.number().nullable().describe("Current installment number if applicable"), totalInstallments: z.number().nullable().describe("Total number of installments if applicable"), currency: z.string().default("BRL").describe("ISO currency code") }); // Extract transaction data from user input export async function extractTransactionFields(userInput: string) { try { const result = await generateObject({ model: openai('gpt-4-turbo'), schema: transactionSchema, prompt: `You are a financial data extraction expert for Finance Mind. Extract structured transaction data from the user's input. INPUT GUIDELINES: - Input may be in any language (primarily Portuguese, English, or Spanish) - Input may contain transaction details in natural language format - Input may include dates in various formats - Input may include currency amounts - Input may include merchant names OUTPUT REQUIREMENTS: Return a structured transaction object following the schema provided. EXAMPLES: Input: "150 para corrida de uber ontem" Output: { "description": "Corrida de Uber", "amount": -150, "date": "2023-09-15T00:00:00.000Z", // assuming yesterday was this date "merchant": "Uber", "category": "TRANSPORTATION", "type": "EXPENSE", "isInstallment": false, "installmentNumber": null, "totalInstallments": null, "currency": "BRL" } Input: "${userInput}"`, temperature: 0.2, // Low temperature for consistent results }); // Check for warnings if (result.warnings && result.warnings.length > 0) { console.warn('Extraction warnings:', result.warnings); } return result; } catch (error) { console.error('Transaction extraction error:', error); throw error; } }
import { generateObject } from 'ai'; import { openai } from '@ai-sdk/openai'; import { z } from 'zod'; // Define the categorization schema const categorizationSchema = z.object({ category: z.string().describe("Predefined transaction category"), type: z.enum(["INCOME", "EXPENSE", "TRANSFER"]).describe("Transaction type"), confidence: z.number().min(0).max(1).describe("Confidence level from 0.0 to 1.0"), merchant: z.string().nullable().describe("Formatted merchant name if available"), isInstallment: z.boolean().describe("Whether this is an installment payment"), installmentNumber: z.number().nullable().describe("Current installment number if detected"), totalInstallments: z.number().nullable().describe("Total number of installments if detected"), isInternalTransfer: z.boolean().describe("Whether this is a transfer between user's own accounts"), tags: z.array(z.string()).describe("Optional tags for further classification") }); // Define tools for enhanced transaction categorization const tools = [ { name: "checkMerchantDatabase", description: "Check for merchant information in our database to aid categorization", parameters: z.object({ merchantName: z.string().describe("The name of the merchant to look up"), }), }, { name: "detectTransferPattern", description: "Detect if a transaction description matches known transfer patterns", parameters: z.object({ description: z.string().describe("The transaction description to analyze"), }), } ]; // Categorize transaction data export async function categorizeTransaction(transactionData) { try { const result = await generateObject({ model: openai('gpt-4-turbo'), schema: categorizationSchema, tools, prompt: `You are a financial transaction categorization expert for Finance Mind. Analyze the transaction details and determine the appropriate category, type, and other metadata. INPUT: ${JSON.stringify(transactionData, null, 2)} CATEGORIES: FOOD_DINING: Restaurants, food delivery, groceries, etc. SHOPPING: Retail purchases, online shopping, etc. TRANSPORTATION: Uber, taxis, public transit, fuel, etc. UTILITIES: Electricity, water, internet, phone, etc. HOUSING: Rent, mortgage, home maintenance, etc. HEALTH_FITNESS: Medical expenses, gym, pharmacy, etc. ENTERTAINMENT: Movies, games, streaming services, etc. TRAVEL: Hotels, flights, vacations, etc. EDUCATION: Tuition, books, courses, etc. PERSONAL_CARE: Haircuts, spa, beauty products, etc. INVESTMENT: Stock purchases, mutual funds, etc. INCOME: Salary, freelance income, etc. TRANSFER: Money transfers, credit card payments, etc. CREDIT_CARD_PAYMENT: Credit card bill payments SUBSCRIPTION: Regular subscription services OTHER: Miscellaneous expenses BRAZILIAN FINANCIAL SYSTEM SPECIFICS: - PIX transactions are transfers and should be categorized as TRANSFER - "Compras parceladas" (installment purchases) should have isInstallment=true - "Boletos" are typically bill payments - IOF is a Brazilian financial transaction tax often seen with foreign currency transactions`, temperature: 0.2, toolChoice: "auto", }); // Check for warnings if (result.warnings && result.warnings.length > 0) { console.warn('Categorization warnings:', result.warnings); } // Log request body for debugging if needed // console.log('Request body:', result.request.body); return result; } catch (error) { console.error('Transaction categorization error:', error); throw error; } }
Model Selection:
openai('gpt-4-turbo')openai('gpt-3.5-turbo')Temperature Settings:
temperature: 0.2Tool Design:
.describe()Error Handling and Debugging:
result.warningsresult.request.bodySchema Design:
Testing and Iteration:
Project Integration:
// src/lib/ai/index.ts import { openai } from '@ai-sdk/openai'; // Configure OpenAI provider export const openaiModel = openai( process.env.OPENAI_MODEL || 'gpt-4-turbo', { apiKey: process.env.OPENAI_API_KEY } ); // Default configuration export const defaultConfig = { temperature: 0.2, maxTokens: 1000, };
Schema Organization:
// src/schemas/transaction-schemas.ts import { z } from 'zod'; export const transactionBaseSchema = z.object({ description: z.string().describe("Transaction description"), amount: z.number().describe("Transaction amount"), // ... other common properties }); export const extractionSchema = transactionBaseSchema.extend({ // Additional extraction fields }); export const categorizationSchema = z.object({ // Categorization specific fields });
Server Actions Integration:
// src/actions/ai/extract-transaction-fields.ts 'use server'; import { generateObject } from 'ai'; import { openaiModel, defaultConfig } from '@/lib/ai'; import { extractionSchema } from '@/schemas/transaction-schemas'; export async function extractTransactionFields(input: string) { // Implementation using Vercel AI SDK }
UI Integration:
'use client'; import { useCompletion } from 'ai/react'; function TransactionInput() { const { completion, input, handleInputChange, handleSubmit, isLoading } = useCompletion({ api: '/api/extract-transaction', }); // UI implementation }
Remember to inspect and handle warnings from the AI SDK calls to ensure proper integration, and implement appropriate error handling throughout your application.