View Agent
Open in Wonderful Platform
Overview
| Field | Value |
|---|---|
| Market | Israel - Banking |
| Vertical | Financial Services |
| Model | GPT-4 Realtime |
| Company | Discount (דיסקונט) |
| Persona | Female, Friendly, Efficient, Smart |
| Language | Hebrew |
| Channel | Voice |
| Architecture | Dynamic skill-based routing |
Skills
Recent Account Transactions - Retrieves and details recent executed current-account transactions (Debit, Credit, or both) within a requested date range (max 3 months back). Active Cards Inquiry - Provides real-time information about all active credit cards associated with the customer’s account. Checks Activity Inquiry - Informs on checks drawn from the account or returned checks, including debit dates, amounts, and detailed reasons for returned items. Salary Payment Check - Specifically identifies and reports the date and amount of salary deposits made to the checking account within the last 3 months. Credit Card Debit Inquiry - Handles all inquiries regarding actual money debited from the account, including past charges, current billing cycles, and upcoming future charges.Key Tools
| Tool | Purpose |
|---|---|
| Recent Account Transactions | Fetches history of account movements (debits/credits) for a specific timeframe. |
| Active Cards Inquiry | Lists all valid credit cards and their basic statuses. |
| Checks Activity Inquiry | Monitors the status of outgoing checks and identifies issues with deposited checks that were returned. |
| Salary Payment Check | Scans account history to locate specific salary-labeled incoming transfers. |
| Credit Card Debit Inquiry | Provides account-level summaries and card-level transaction details for past and future charges. |
| Credit Limit Inquiry | Retrieves the current credit frame and available balance for the customer’s cards. |
Prompting Techniques
Dynamic Skill Architecture: Uses a lean Base Prompt to minimize token usage, relying on a “Dynamic Skill” logic where each specific skill contains its own dedicated prompt, activated only when the intent is matched. OpenAI Cookbook Alignment: Instructions are structured according to OpenAI best practices for high-performance Realtime models, focusing on clear tool definitions and context preservation. TTS-Ready Formatting: Strict formatting rules ensure numbers, currencies, and dates are converted into natural spoken Hebrew words, avoiding symbols that might confuse Text-to-Speech engines. Example Prompt Snippet: “When saying currency values (e.g., balance, charges), you must separate components using commas - between thousands, hundreds, shekels, and agorot. Example: for 999,694.82, say: ‘תשע מאות תשעים ותשע אלף, שש מאות תשעים וארבעה שקלים, ושמונים ושתיים אגורות’.”Lessons Learned
What worked- Centralized Base Rules: Placing generic rules (gender, currency formatting, TTS standards) in the Base Prompt ensured consistency across all 28 skills without duplication.
- Global Tooling: Implementation of global tools (like satisfaction surveys and human banker transfers) that are accessible regardless of the active skill.
- Semantic Naming: Precise naming and descriptions of skills and tools are the most significant factors for a dynamic agent, as they dictate the accuracy of the skill-switching logic.
- API Complexity: Mapping 21 different API interfaces into specialized tools for a multi-skill agent (28 skills) was a significant architectural challenge.
- CLI Development: Transitioning to code-based agent management via the CLI platform required a learning curve.
- Skill Mapping: Managing a large volume of skills required creating a strict mapping of which tools belong to which intent to avoid logic overlaps and confusion.