AI That is Credit Smart III: Smart Prompting for CreditCompanion™

The Logic Behind Retrieval Augmented Generation (RAG) Models Like CreditCompanion^TM

To understand what makes a good prompt for CreditCompanion^TM, it helps to look at what happens behind the scenes. CreditCompanion^TM uses Retrieval Augmented Generation (RAG), an advanced AI method that combines relevant information search with insightful answer generation.

How CreditCompanion^TM Processes Your Question

When you enter a question, the following steps occur in the backend: (Chart 1)

1. Text Conversion: CreditCompanion^TM first converts your text into a mathematical format that AI models can process.
2. Intent Analysis: The system then analyzes the question to infer your intent- whether you are conducting entity analysis, looking up criteria, performing macro research, or running a data query etc.
3. Smart Routing: Based on what it learns from your question, the system routes it to the most appropriate process, including criteria search, definition lookups, cross-article summaries, entity-level research and updates, and macroeconomic and industry analysis.
4. Response Generation: These relevant chunks are brought together and passed through a generative AI model, which creates a customized response based on your question via token generation.

This process ensures that CreditCompanion^TM delivers reliable and context-aware answers by blending powerful retrieval with intelligent generation.

Chart 1:  Illustration of a custom-built RAG solution to retrieve and summarize credit risk related content and research

Why Clear Prompts Matter

How RAG works has a direct impact on what makes a good prompt.  Large language models infer your intent and the type of content you need based solely on your question and how it is mathematically represented.  Since the model cannot read your mind, providing clear and specific prompts helps ensure you receive the most accurate and relevant responses.

Eight Essential Tips for Effective Prompting

Tip 1: Be Clear in Your Intent

Since CreditCompanion^TM is custom designed for credit risk workflows, all responses are tailored toward credit-related use cases at the system prompt level.

For example, if you asked CreditCompanion^TM "What is the weather today?" (Chart 2) it will not provide a response. Instead, it will point you toward the different use cases it supports:

• Macro Research
• Entity Research
• All Research
• Criteria
• Definitions
• Data Query

For instance:

• If you request a company's ratings history and use the Data Query option, CreditCompanion^TM will query the underlying database for the information you need and will not be extracting potentially stale information from articles.
• If your question requires searching across multiple research articles (Macro Research, Entity Research, All Research), the system filters articles available in RatingsDirect^® by industry, country, and entity identifier. It then divides the text into smaller chunks and retrieves the sections that are most relevant to your query.

Clarifying your intent at the beginning of the prompt can immensely improve the quality of the responses as it directs the large language model to the right process for your use case.

Chart 2: Example of CreditCompanion’s response to a prompt which will generate a response with general purpose chatbots, but lies outside the scope of CreditCompanion^TM

How to clarify your intent: Use the "#" symbol in your prompt, followed by your intent. Typing "#" brings up various options. For example, if you want to focus on Macro Research, simply select it from the menu. (Chart 3)

Chart 3: Using “#” at the start of the prompt to clarify user intent in CreditCompanion^TM

Tip 2: Be Clear About Entities and Geographies

Most chatbots rely purely on the text you type to filter content, but CreditCompanion^TM has additional functionality to help you be more precise. Use the "@" sign to specify exactly which entity, geography, or even article you are interested in searching.

Using the "@" sign brings up a menu where you can specify:

• Country: For example, @Singapore
• Entity: For example, @Intel Corp
• Specific Article: You can even restrict responses to a single article

Example Comparison:

• General Search: When a user asks about Singapore as a country (i.e. using @Singapore) and risks from the war in the Middle East, CreditCompanion^TM will prefer to summarize content across articles that tag  Singapore as a geography. (Chart 4).

Chart 4: Using “@” within the prompt to clearly specify geography locations being examined

• Focused Search: Using "@<article name>", selecting S&P Ratings Research and selecting Full Analysis article on Singapore's sovereign rating, published in May 2026, will focus purely on responses from that specific article and on the sovereign entity (Chart 5)

Chart 5: Using “@” within the prompt to clearly specify article to be summarised

Tip 3: If your Question is Unrelated to the Previous Ones, Open a New Chat

Large language models are highly sensitive to context.  Even small changes in the prompts you entered previously, shift the probabilities of the next generated word. To ensure the most accurate responses, start a new chat for each new question (i.e. clear the context window). Avoid passing previous chat history, timestamps, or user session data into your next prompt.

Tip 4: Provide a Blueprint of What the Output Should Look Like

For more elaborate prompt engineering, you may want to write out exact input-output pairs inside your prompt template.  However, even for CreditCompanion^TM, it can be helpful to anchor the logic of what you want by clearly outlining the structure of your desired response.

For example, if you are interested in focusing on the government's fiscal policy measures, industries most affected, and mitigants, list them clearly in point format. Through pattern recognition, the large language model from CreditCompanion^TM understands that you are looking for replies in a particular structure and follows suit.

Example prompt structure to give you the specific aspects that you are searching for within a broad topic:

"Please provide information on: 1. Government's fiscal policy measures.  2. Industries most affected. 3. Risk Mitigants"

This approach anchors the model's pattern-matching mechanics, making it heavily bias toward copying the structure of your request. (Chart 6)

Chart 6: Example of a prompt that guides CreditCompanion^TM to provide output in a structured format

Tip 5: Use Constraints and Boundaries

Set word limits, time periods, or scope restrictions to focus your results. Setting a shorter timeframe and a longer word count can give you a more detailed analysis. Examples:

• "Focus only on developments in the last 12 months"
• "Limit response to 200 words"
• “What is the latest credit risk outlook for @<company name>. Use only research published from January 2026 to present.”

Tip 6: Ask for Specific Formats and Rank Ordering of Responses

Requesting bullet points, comparisons, or pros/cons lists can make information easier to digest.

Example (as a follow up question on Intel Corp.): "Can you list the top 5 risks in order of severity?"

Tip 7: Ask Follow-up Questions Within the Same Context Window

In a similar way that you would clarify with a person if a response was unclear or not what you wanted to hear, you can similarly ask CreditCompanion^TM follow-up questions to narrow down or expand on the information provided. Examples:

• "Can you elaborate on the fiscal policy aspect?"
• "Focus more on the banking sector"

Tip 8: Avoid Ambiguous Language

Providing specific rather than vague language will remove the guesswork from the large language model (i.e. it decreases the noise in the numerical presentation of your question).  This increases the chances of getting the responses you are looking for.

Use concrete terms like dates, specific metrics, or named entities instead of vague terms such as "recent," "some," or "various." If there is a specific angle of risk that you are concerned about, you can specify these as well (e.g., supply chain risks).

Instead of: "What are recent developments?"
Try: "What developments occurred between January and May 2026?"

Instead of: "What are some risks?"
Try: "What are the top supply chain risks affecting the technology sector?"

Conclusion

By understanding how CreditCompanion's RAG model works and applying these eight smart prompting tips, you can unlock more accurate, relevant, and actionable insights. Clear intent, precise entity selection, isolated requests, and structured outputs are the keys to getting the most out of your CreditCompanion^TM experience.

Disclosures for this article from S&P Global Market Intelligence:
https://www.spglobal.com/marketintelligence/en/legal/disclosures#sp-global-market-intelligence