Prompting Guide Unique FinanceGPT Chat External Knowledge & Recording
- 1 Prompting Techniques
- 1.1 1. Clear, Direct Statements
- 1.2 2. Language of the Report (Only for Recording Platform)
- 1.3 3. What to Do Rather Than What Not to Do
- 1.4 4. Persona Pattern
- 1.5 5. Audience Persona Pattern
- 1.6 6. Context of the Recording
- 1.7 7. Defining the Scope and Depth
- 1.8 8. Using the Right Keywords
- 1.9 9. Setting the Tone or Style
- 1.10 10. Requesting a Specific Format
- 1.11 11. Template Format
- 1.12 12. Multiple Choice Prompts
- 1.13 13. Zero-Shot Prompts
- 1.14 14. Providing Examples
- 1.15 15. Cognitive Verification
Prompting Techniques
Prompting techniques are crucial when using LLM’s because they directly influence how effectively the AI understands and responds to user requests.
1. Clear, Direct Statements
Explanation:
Clear and direct statements in prompts help avoid misunderstandings and enhance the efficiency of communication between humans and AI. A well-formulated prompt should be unequivocal and unambiguous, so that the AI understands exactly what is expected of it.
Example Prompt:
Example Recording:
"What is mentioned in the transcript about the hobbies, goals, and desires of the customer?"
is better than
"What is mentioned in the transcript about the customer?"
Example External Knowledge:
“Tell me which three factors most significantly impact the profitability of banks and insurance companies, and how they influence their financial performance”
is better than
“Tell me something about banking and insurance?”
2. Language of the Report (Only for Recording Platform)
Explanation:
If you would like the report in the same language as the recording, you can select this in the settings. However, if you want the report to always be in the same language, you must specify this at the beginning of the prompt.
Example Prompt:
“Force to write in German. Summarize the transcript.”
3. What to Do Rather Than What Not to Do
Explanation:
It is more effective in prompts to directly state what should be done rather than describing what should NOT be done. This leads to a more direct and productive response from the AI. The AI is not good to interpret the words NOT and DON’T.
Example Prompt:
Example 1:
"What is mentioned in the transcript about the hobbies, goals, and desires of the customer?"
is better than
“What is mentioned in the transcript about the customer but don’t tell me the birthday and do not mention the address. ”
Example 2:
“What is mentioned in the transcript about the customer, avoid telling me the birthday and refrain from mentioning the address”
is better than
“What is mentioned in the transcript about the customer but don’t tell me the birthday and do not mention the address. ”
4. Persona Pattern
Explanation:
By assigning a specific persona, the prompt can generate specific and context-dependent responses that match the character or role of the persona. This is particularly useful in scenarios where the perspective or tone of the response is important.
Example Prompt:
Example Recording:
"Act as a skeptical scientist and summarize the transcript."
Example External Knowledge:
"Act as a skeptical scientist and assess the reliability of current research on artificial intelligence."
5. Audience Persona Pattern
Explanation:
The "Audience Persona Pattern" refers to the technique of considering a specific persona when constructing prompts. This method helps align the AI's responses better with the needs, expectations, and context of the target audience. A persona can be based on various characteristics such as age, occupation, interests, or behaviors. By applying this pattern, the interaction between the user and the AI can be personalized and more relevant.
Example Prompt:
Example 1:
"Assume I am a new bank clerk with limited experience in using financial software. Please summarize the transcript"
Example 2:
"Assume I am a new bank clerk with limited experience in using financial software. Please explain how I can use a banking system to process customer deposits. Include step-by-step instructions starting from logging into the system to finalizing the transaction."
6. Context of the Recording
Explanation:
The context of a conversation refers to the specific circumstances under which the conversation took place. This can include the topic of the conversation, the participants involved, the time and place of the conversation, and the purpose of the discussion.
A clear understanding of the context is crucial for correctly interpreting the contents of the transcript and filtering out relevant information. The topics discussed in a banking transcript can vary from general inquiries to specific product information to complaints. It is important to recognize the main topics of a conversation to properly categorize the discussions and to understand which information is significant for the bank and the client.
Each conversation can have specifics that are important for understanding the full context. These can include special requests from the client, unusual incidents during the conversation, or a particular emotional mood. These details are often crucial for interpreting customer needs and expectations.
Example Prompt:
"Analyze the attached transcript of a consultation between a customer advisor and a client of Bank XYZ, which took place in Munich. The main purpose of the conversation was to advise on real estate financing. Please identify the main topics discussed, as well as any concerns or special requests from the client."
7. Defining the Scope and Depth
Explanation:
When prompting, it is crucial to clearly define the scope and depth of the desired response. This helps the AI understand how comprehensive or detailed the answer should be. A well-defined scope avoids overly general or specific answers that do not meet expectations.
Example Prompt:
Example 1:
"Explain out of the transcript what was said of the basic principles of investing with a focus on EFT’s for an audience that is new in this area."
Example 2:
"Explain the fundamental principles of risk management in banking, with a focus on credit risk assessment, for an audience familiar with basic financial concepts"
8. Using the Right Keywords
Explanation:
Keywords are essential for directing the AI response's direction and focus. Selecting precise keywords helps the AI better understand the topic and filter relevant information more effectively.
Example Prompt:
Example Recording:
"Analyze the transcript of the conversation between the customer advisor and the customer regarding investments describe the most important topics particularly, Return on Investment (ROI), Risk Assessment, Market Analysis”
Example External Knowledge:
"Describe the latest advancements in artificial intelligence, particularly machine learning, neural networks, and natural language processing."
9. Setting the Tone or Style
Explanation:
The tone or style of a prompt guides the AI on how the answer should be formulated. This can range from formal and informative to casual and entertaining, depending on the target audience and purpose of the text.
Example Prompt 1:
"Write a short, humorous summary about the above transcript from an internal meeting."
Example Prompt 2:
Example Recording:
"Write a short summary about the above transcript from an internal meeting. Copy the tone and style from the following example:
Bank Name: Unique Bank
Date: July 1, 2024
Customer Name: John A. Doe
Customer ID: 987654321
Account Number: 1234567890
Type of Account: Checking
Customer Address:
Banhofstrasse 1
8001 Zürich
Contact Information:
Phone: +41 76 000 00 00
Email: johndoe@example.com
Date of Birth: August 15, 1980
Nationality: Suisse
Occupation: Software Engineer
Employer: Tech Innovations LLC
Source of Funds: Salary
Annual Income: 125’000 CHF"
Example External Knowledge:
“Write a short blog post about the challenges of working from home. Copy the tone and style from the following example:
Earlier this week, Unique hosted a dynamic event that brought together top-notch professionals from the banking and wealth management world. Attendees included representatives from prestigious institutions such as LGT, SIX Group, Vontobel, Zuger Kantonalbank, Lombard Odier, and many others. The spotlight of the event was a presentation by Julius Baer, a leading wealth management company, showcasing their use of a robust tech stack combining on-prem models with advanced LLMs like GPT-4.”
10. Requesting a Specific Format
Explanation:
Requesting a specific format helps obtain the AI's response in a particular structure. This can be an essay, a report, a list, a dialogue, or any other form suitable for presenting the information.
Example Prompt:
Example Recording:
"Summarize the above transcript in a list with bullet points."
Example External Knowledge:
"Create a detailed list of the top 10 SEO strategies for 2024, including the benefits and potential drawbacks of each strategy."
11. Template Format
Explanation:
Using a template format in prompting ensures that the AI's output adheres to a predefined structure, which can be particularly useful for generating consistent and organized responses.
Example Prompt:
“From the above transcript, write an email to a colleague.
The structure for the email should look as follows:
Hello Everyone,
Topic 1
Summary in bullet points
Topic 2
Summary in bullet points
etc.
– – – – –
Customer Request:
Summary in bullet points.
Customer's phone number or email:
Thank you and best regards”
12. Multiple Choice Prompts
Explanation:
Multiple choice prompts present the AI with a question and several potential answers from which it must choose or recommend the most appropriate one.
Example Prompt:
Example Recording:
"Take the information from the above transcript. Given a customer's interest in securing a mortgage with the lowest possible interest rate, which of the following options should they consider? A) Fixed-rate mortgage, B) Adjustable-rate mortgage, C) Interest-only mortgage. Explain the benefits and drawbacks of each option based on the current economic conditions."
Example External Knowledge:
"Given a customer's interest in securing a mortgage with the lowest possible interest rate, which of the following options should they consider? A) Fixed-rate mortgage, B) Adjustable-rate mortgage, C) Interest-only mortgage. Explain the benefits and drawbacks of each option based on the current economic conditions."
13. Zero-Shot Prompts
Explanation:
Zero-Shot Prompts refer to the ability of an AI model to understand and handle a task without prior specific training or examples. This type of prompting is used when no specific training data are available, but an immediate response from the AI is expected. This is particularly useful in situations where quick and flexible solutions are required.
Example Prompt
Example Recording:
"Please analyze the following transcript and identify all important financial information, including risk factors, investment recommendations, and customer preferences. Summarize this information in a clear list."
Example External Knowledge:
"Explain the potential impacts of raising the prime rate by 0.5% on the demand for mortgage loans and business loans."
14. Providing Examples
Explanation:
Providing examples in a prompt can help the AI better understand the request and improve the quality of the response. This is especially important when specific formats or detailed information are required.
Example Prompt:
"Make a CRM entry from the transcript. A good CRM entry should contain the following information: customer name, contact information, Company, Industry, Role, Purpose of Contact. Create a CRM entry from the Transcript.
CRM Entry Example
Customer Name: John Doe
Contact Information:
Phone: +1 234 567 8910
Email: johndoe@example.com
Address: 1234 Maple Drive, Springfield, IL, 62701
Company: Doe Enterprises
Industry: Manufacturing
Role: Procurement Manager
Purpose of Contact:
Discussed John's ongoing needs for upgraded manufacturing equipment.
Followed up on the proposal sent on May 30, 2024.
“
15. Cognitive Verification
Explanation:
Cognitive Verification refers to the process of breaking down a complex question into smaller, manageable parts. This makes it easier for the LLM to provide detailed and accurate responses. This method encourages the generation of additional questions that refine understanding and collect comprehensive details. The responses to these individual questions are then combined to formulate a final answer.
Example Prompt:
Example Recording:
Situation:
Suppose you want to understand from a transcript of a banking conference what new compliance regulations were introduced and what impacts these might have on business processes.
"Given a transcript of a banking conference, break down the discussion on new compliance regulations into the following parts:
What specific compliance regulations were discussed during the conference?
What reasons were given for the introduction of these new regulations?
What specific changes in business processes are required by these regulations?
What concerns or challenges were raised by participants regarding the new regulations?
How were these concerns addressed?
Combine the answers into a comprehensive report on the new compliance regulations and their potential impacts on the bank."
Example External Knowledge:
"Imagine you are an LLM tasked with helping me, a bank administrator, understand how to reconcile discrepancies in transaction reports. First, could you list what types of discrepancies commonly occur in banking transaction reports? Next, what steps should be followed to investigate these discrepancies? Finally, how should the findings be documented and reported?"
For further techniques also see:
https://jmservera.github.io/miscdemos/prompt-engineering#additional-resources
https://www.coursera.org/learn/prompt-engineering
Author | @Cornelia Hauri |
---|---|
Version | V 1.0 (Date: 11.07.2024) |
© 2024 Unique AG. All rights reserved. Privacy Policy – Terms of Service