How can I reliably compute relative dates (e.g., 'this Friday') using an LLM?

Below is an approach that many have had success with:

1. Separation of Concerns:
   Instead of relying on the LLM to perform date arithmetic reliably, you can use the LLM to extract the relative date expression from user input and then pass that phrase to a dedicated date-parsing library. This way, the LLM focuses on understanding natural language, and the library handles date computation.

2. Using Date Parsing Libraries:
   Libraries such as dateparser for Python (or libraries like Chrono for JavaScript) are designed to translate phrases like “next Thursday” or “2 days from now” into dates accurately. For example, in Python you might use:

   PYTHONCopied
   import datetime
   import dateparser
   
   def get_relative_date(relative_expression, base_date=None):
       if base_date is None:
           base_date = datetime.datetime.now()
       settings = {
           'RELATIVE_BASE': base_date,
           'PREFER_DATES_FROM': 'future'  # This setting helps with phrases like 'Friday'
       }
       parsed_date = dateparser.parse(relative_expression, settings=settings)
       return parsed_date
   
   # Example usage:
   base = datetime.datetime(2025, 1, 28)
   relative_expression = "this Friday"
   result = get_relative_date(relative_expression, base_date=base)
   print(result.strftime("%Y-%m-%d"))  # Expected output: 2025-01-31

   This isolates the date calculation in a robust library while the LLM can confirm or extract the right natural language fragment.

3. Combining LLM and Date Libraries Carefully:

   • Extraction: Let the LLM extract relevant phrases indicating relative dates from the text.
   • Validation: Once extracted, validate or adjust the expression if needed before feeding it into the parser.
   • Fallbacks: Consider a fallback strategy if the LLM output is ambiguous. For instance, if an expression like “next Friday” is misinterpreted, you might add context such as the current day of the week to help the parser choose the correct date.

4. Unit Testing:
   Given that different contexts or phrasings might result in edge cases, ensure you have good test coverage. Create tests with known dates (using fixed base dates) and relative expressions to verify the correctness of the output.

5. Iterative Refinement:
   It might be helpful to tweak your prompt for date extraction from the LLM so it consistently returns expressions in a predictable format. A structured prompt may yield better extraction results.

This hybrid approach lets you leverage the strengths of natural language understanding from LLMs while relying on proven date arithmetic logic from specialized libraries, leading to more consistent and maintainable behavior in your LLM application.