API Prompt for gpt-3.5-turbo-16k

It appears the solution to this was to send the prompt formatted as XML code:

      xmlPrompt : 
            <text>What is Mark's IP address based on the conversation?</text>
          <!-- Root element containing all context documents -->
          <!-- Context Document 1 -->
          <contextDocument id="1">
            <!-- Title of the document -->
            <title description="The title of the document.">my-queries-export.csv</title>
            <!-- Main content of the document -->
            <content description="Main content of the document.">
            <!-- Summary of the document -->
            <summary description="Brief summary of the document."></summary>
            <!-- Event date of the document -->
            <eventDate description="Date of the event or document creation."></eventDate>
            <!-- Groups associated with the document -->
            <groups description="Groups or categories associated with the document.">Documentation, Development, D9 Development, AI</groups>
            <!-- Taxonomy or tags of the document -->
            <taxonomy description="Taxonomy or tags for categorization."></taxonomy>
            <!-- URL of the original document -->
            <url description="URL of the original document.">https://master1and1-9.schoolboard.net/system/files/my-queries-export.csv</url>
            <!-- Questions that this document answers -->
            <questions description="Questions that this document answers."></questions>
          <!-- Context Document 2 -->
          <contextDocument id="2">

As I initially surmised, the problem, besides gpt-3.5-turbo-16k being considerably less efficient at reading text than the gpt-4 family, was the format of my submitted prompts. This one tweak has resulted in much better (though far inferior to it’s gpt-4 cousins) responses:

A step in the right direction, at the very least.

1 Like