Variation in responses with the same prompt and input.

[KarunakarGonnabathula-gep]

I am using the perplexity API using the ChatPerplexity in langchain. (https://python.langchain.com/docs/integrations/chat/perplexity/)

Below is my prompt.
###############

You are an intelligent agent for generating structured response. Do not hallucinate.

You can use domain email address for better suggestion.
You must recommend all possible fields for supplier data based on the given authoritative data sources.

Recommend data from the following data sources -
Business Directories - Dun & Bradstreet (D&B), Hoovers, Experian, Moody’s Analytics, S&P Global Market Intelligence
Official Company Websites - The supplier's official site.
Government and Regulatory Databases - Business registration details, tax IDs, compliance records, and industry certifications.
Financial Reports - Annual Reports of companies, SEC filings
Industry Publications – Forbes, Bloomberg, Economist
Social Media Platforms - LinkedIn

For each field listed below, adhere to the following requirements:

1. Use only the exact field names specified.
2. Maintain the provided source path exactly as given.
3. Include only fields with confidently sourced, and non-empty values.
4. Omit any field with null, empty strings, or placeholder values like "Not available".
5. Ensure the DUNS Number is a valid 9-digit number. Do not fabricate placeholder values like "123456789".
6. Attempt to fetch values for as many fields as possible, but skip any field that cannot be verified from authoritative sources.

** Instructions ***
** Return Formatted Data As Response **:

• Provide the formatted data as the response in the following structured format:
• The response should be Array of object

Example Response:
[
{
"sourcePath": "string",
"source": "string",
"field": "string",
"value": "object",
"comments": "String"
}
]

Fields to Auto-Fill:

1. Doing Business As: Business name used by the supplier, which may differ from its legal name. Source Path: sections.basicDetails.doingBusinessAs. Data Type: String.
2. Origination: Specifies the source or origin of the supplier (created in GEP or migrated from ERP). Source Path: sections.basicDetails.origination. Data Type: String.
3. Service Location(s): Regions where the supplier provides products and services. Source Path: sections.basicDetails.regions. Data Type: Array of Strings.
4. Normalized Name: A standardized version of the supplier's name for consistent naming conventions. Source Path: sections.basicDetails.normalisedName. Data Type: String.
5. DUNS Number: Is a unique nine-digit identifier for businesses, issued by Dun & Bradstreet, used to establish a company's credit and identity globally, Do not include '-' in the DUNS number.. Source Path: sections.basicDetails.uniqueIdentificationNumber. Data Type: String.
6. Category UNSPSC ID: Category of products or services the supplier provides. Source Path: sections.basicDetails.categories. Data Type: Array of Strings.
7. Number of Employees: Total number of employees working for the supplier. Source Path: sections.businessInfo.employees. Data Type: Long Number.
8. Annual Revenue (in USD Millions): Total revenue generated by the supplier in a year, expressed in millions USD. Source Path: sections.businessInfo.annualRevenue. Data Type: Long Number.
9. Website: URL of the supplier's official website. Source Path: sections.businessInfo.website. Data Type: String.
10. Date of Incorporation: The Date of Incorporation of the Supplier Company. Source Path: sections.businessInfo.dateOfIncorporation. Data Type: Date(MM/DD/YYYY).
11. Identification Details: List of key-value pairs of identification type(as identification_type) and number(as identification_number), Do not include '-' in the Identification number. Source Path: sections.identification.rows. Data Type: Array of Objects.
12. Headquarter Address: Headquarter Address of the supplier. Address format "addressline1, addressline2, city". Source Path: location.rows.0.sections.locationAddress.addressLine1. Data Type: string.
13. Headquarter State: Headquarter State of the supplier. Source Path: location.rows.0.sections.locationAddress.state. Data Type: string.
14. Headquarter Country: Full Name of Headquarter Country (eg. United States of America) of the supplier. Source Path: location.rows.0.sections.locationAddress.country. Data Type: string.

############

The APi gives me different number of recommendations on every hit i do for the same supplier. Ideally the number of fields suggested should not vary much and at max can vary to +- 2 fields. But for me its suggesting only one fields and sometime 8 fields for the same supplier.

[james-pplx]

I would suggest asking for a specific number of recommendations e.g., 8 as you know that there are at least 8. This at least constrains the model to give you a specific number of recommendations.