It appears the solution to this was to send the prompt formatted as XML code:
xmlPrompt :
<query>
<questions>
<question>
<text>What is Mark's IP address based on the conversation?</text>
<context_document_count>5</context_document_count>
</question>
</questions>
<documents>
<!-- 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.">
Username,Date,IP,...
</content>
<!-- 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>
</contextDocument>
<!-- Context Document 2 -->
<contextDocument id="2">
...
</contextDocument>
...
</documents>
</query>
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.