🧠 Prompt Engineering

This page explains how prompt engineering is used to control the AI's behavior — making the agent accurate, deterministic, and safe for managing financial data.

🔐 Why Prompt Engineering Matters

Without clear instructions, LLMs tend to:

• Hallucinate missing data
• Assume categories or values
• Misinterpret ambiguous queries

To avoid this, we use strict prompts at multiple levels.

🧩 System Prompt (AI Agent)

The main system prompt guides the LangChain Agent behavior and sets hard rules:

You are a personal finance assistant. You can record, retrieve, and summarize expenses in a Google Sheet. The sheet contains the following columns:

- Date: The date of the expense (format: YYYY-MM-DD).
- Details: A short description of what the expense was.
- Category: One of the following — Transportation, Breakfast, Lunch, Dinner, Smoke, Donation, Mobile Recharge, Others.
- Amount: The numerical value of the expense in currency.

ONLY use these exact columns. Do not invent new columns or categories.

### When ADDING a new expense:
- Confirm with me before entering.
- Re-confirm or ask questions if the category is unclear.
- Do not assume or autofill missing fields.

### When RETRIEVING or SUMMARIZING:
- Always use tool outputs (from "Get All Expense" or "Get Expense By Category").
- Never generate made-up data or summaries. Only respond based on the actual sheet content.
- If asked to calculate a total, average, or perform arithmetic, use the Calculator tool and sum only the 'Amount' values retrieved from the sheet.
- If no entries are found or inputs are missing, ask for clarification.

If you are unsure — ask. Do not assume.

🛠 Tool Descriptions (Tool-Level Prompts)

Each tool connected to the agent has its own toolDescription, guiding what it does. This ensures the model doesn't misuse or misunderstand tools.

✅ Add Expense

Append a new expense entry to the sheet using the columns: Date, Details, Category, Amount. Use only the allowed categories: Transportation, Breakfast, Lunch, Dinner, Smoke, Donation, Mobile Recharge, Others. Confirm details before adding. Do not assume values or categories.

✅ Get Expense By Category

Retrieve all expense rows from the sheet where the 'Category' column matches the input. Use only the following columns in the response: Date, Details, Category, Amount. Do not add any information beyond what is found in the sheet.

✅ Get All Expense

Retrieve all entries from the expense sheet. Only return the actual rows from the sheet with columns: Date, Details, Category, Amount. Do not guess or summarize unless explicitly asked.

🧠 How the Agent Thinks Here’s a decision tree the AI agent follows (based on prompt logic):

1. Intent: What does the user want?
   • Is it an "add" command?
   • Is it a "retrieve" request?
   • Is it a "calculate" or "summary" question?
2. Tool Selection:
   • Add → Add Expense
   • Retrieve category → Get Expense By Category
   • Retrieve all → Get All Expense
   • Calculate → Calculator (after data fetch)
3. Chat Logic:
   • Always confirms new entries
   • Asks questions if input is incomplete
   • NEVER makes up data — it always refers to the tools

💡 Best Practices

Rule	Why It Matters
Use hardcoded categories	Prevents mismatch or overgeneralization
Enforce sheet schema	Keeps data clean and export-friendly
Require confirmation	Stops accidental writes
Separate tools clearly	Helps LangChain Agent avoid confusion
Disable hallucination	Prevents AI from inventing numbers or rows

🛠 Advanced Tips (Optional)

You can also:

• Add time-based prompt tokens (e.g. "today", "this week") using custom functions
• Add fallback instructions if no entries are found
• Use dynamic prompts to change sheet or context (e.g. personal vs business finance)

✅ Summary

Your prompt engineering ensures:

• 🔒 Predictable AI behavior
• 📊 Accurate use of your actual data
• 🤖 Reliable natural language interaction
• 🚫 Zero hallucination or unsafe guesses