I’m using a prompt setup where the chatbot is configured to respond to specific commands (e.g., /upload, /start, /link). The /upload command should trigger an automatic analysis of a document upon upload, with a direct response containing insights from the file.
After adding two documents to the chatbot’s knowledge base—a list of commands and a compilation of 30-40 newsletters as examples—I’ve encountered an issue: whenever /upload is used, the chatbot often fails to respond with the document analysis. Instead, it either repeats phrases (likely from the newsletters), provides random business insights, or misinterprets the command. This may stem from the informal, narrative tone within the newsletters, which seems to confuse the chatbot’s interpretation.
Key Findings (with the help of ChatGPT after analizyng the Newsletter document):
- The chatbot might be mistaking narrative phrases as instructions, causing it to respond inappropriately to /upload and other commands.
- Reference sections in the newsletters seem to interfere with the intended command responses.
When the files in the Knowledge Base are off, the /upload command works.
I’d like the /upload command to apply solely to the user-uploaded file, not to the existing documents containing narrative newsletter content.
Any tips on how to prevent the chatbot from confusing reference material in the newsletters with actionable commands would be appreciated.
Alternatively, should I simplify the command structure or modify the newsletter content itself to avoid this interference?
Thank you all!