I hope I have selected the correct option for this help request. I have built this agent and it’s function is to extract the documents that are needed from the borrower and create an email requesting those items.
I need help with a couple issues:
I have used ChatGPT to help fine tune the output results. I gave the LLM multiple examples of the source docs for context, as well as the template I wanted the output to be formatted in, when the run is complete. I have noticed the wording in the output is slightly different from the phrasing I gave the LLM for output. I have the temperature set at 0.0 for the output. (So how do I dial this in, to get my wording in the draft email, I’m going for consistency, so that other users can use the agent and get predictable draft email responses, with very little editing?)
How do I modify the Use input block to accept multiple pdf’s to the agent? I changed the setting to allow more than one pdf, but it doesn’t see to allow more than one pdf to be uploaded at a time.
I have vibe coded the user input page and noticed that in the sections the user has to complete, MS is saving the last input data. (i.e. if I added my name in that block, it seems to be saved as an option the next time I go to complete the block.) The issue is if I used the saved data from the previous runs, the form won’t allow my to click next and start the run. (So how do I get the agent to allow the saved info or make the agent clear the user input blocks after each run?)
Thanks for the post! Let’s go over your questions one by one.
Using markdown formatting and XML tags can help improve output consistency. Also, make sure you’re only referencing variables that are defined earlier in the workflow. If a variable isn’t defined, the model won’t receive its value, since all {{variables}} in curly braces are resolved before being sent to the model.
You can also try using different AI Models to see which one gives the most consistent results. You can test multiple models with the same prompt using the Profiler tab in your Agent’s IDE.
If most of the email stays the same and only certain words or phrases change, you can set the Generate Text block’s output to JSON, then use a Display Content block with an email template and pull in specific JSON key values where needed.
If you have a native interface selected in the User Input block, please make sure to click the Publish button in the top right corner of the IDE so the change takes effect.
If you’d like to let users upload multiple files in the Custom Interface, you can vibe code the Custom interface to allow that, and then use a Run Workflow block to trigger a Sub-workflow with an Extract Text from file or OCR Image block. That way, the User Interface block simply uploads all the files, then you can pass those file links to other blocks to perform extraction.
Could you please record this behavior using the link below?