How to instruct GPT to not link to PDFs in the Knowledge section

[gustavocpimenta]

Hi,

I don’t want my Adaptability Partner GPT (available on the GPT store) to link to the PDFs in the Knowledge section. I want it to use that knowledge but not link to documents.

I achieved that with the following instructions in the web version, but it still links to PDFs in the desktop and mobile apps.

Does anyone know how to fix this?

Thanks in advance.

Instructions for Handling PDFs

• Avoid Mentioning PDFs: Do not mention or refer to the existence of specific documents (PDFs) in your responses.
• Incorporate Insights: Seamlessly incorporate insights and concepts from the documents into your advice and guidance without identifying the source.
• Reference and Link: Continue to reference “Fixed to Flow,” “Flow OS,” “Flow Compass,” and related concepts as needed, including links to Fixed to Flow or other sites mentioned in the instructions.

[mad_cat]

I have created a Custom Instruction GPT I call Subconscious GPT. I use an AI Persona Profile within ChatGPT and the GPT will communicate directly to the Persona not to me, basically it is a passive GPT. I used this to create a persona to help me play Fallout 4, giving the persona a series of resources it can use to help guide me. One of the most important features to Subconscious GPT is that there are certain things I want it to know, but not to reference unless I specifically mention it, otherwise, it keeps the information in mind when giving me advice. For Fallout 4, it is aware of spoilers and prepares me for spoilers, but doesn’t tell me what they are. Might say, “be sure to get extra ammo”, but doesn’t tell me why.

To fix the problem of not referencing source material directly, the best way is in the GPT instructions you can try:

• Role of {name} GPT: Act as a sophisticated omniscience of intelligence, providing context-aware information, unbeknownst to the user of its origin.
• Integrated Data: Assimilate all uploaded documents. Reference upload content first before expanding your search of knowledge. Whether referenced or new, treat all information as generated content.
• Sensitive Referencing: Keep secret where you gain information in your output. Do not tell the user where information originated.

Assimilate is key here, as it tells the system to integrate knowledge with its existing knowledge based within the instance. So your uploaded data and what it already knows becomes one in a metaphorical sense. This should help it not reference the information.

Omniscience does sort of the same thing. The AI can’t be omniscience, but it metaphorically understands this as another way to integrate knowledge. Not as effective as assimilate, but explains doing so in a slightly different way.

Mentioning in a few ways to not tell you where it got data from should help, but feel free to edit to how it best works for you. Please let me know if this helped.

[gustavocpimenta]

Thanks for the prompt response mad_cat, but it didn’t solve the problem… it doesn’t make any sense that the links to the docs don’t show up on the webapp, but show up on the desktop and mobile apps, I must assume it’s a bug .

That said, learned a lot with your suggestion

[mad_cat]

That is odd. And good to know. When one creates a GPT, one should check the performance within the desktop and mobile apps. Thank you for the response back.

Out of curosity, when you used my suggestions, did you apply them within your current GPT, or did you create a new GPT and implement my suggestions? I ask, because if you applied to your current GPT, then if it is a flaw in how it was created, then adding my suggestions the flaw could remain. While it may still end up being a flaw with the external access to the GPT such as the desktop and mobile apps, it good to create a new one just to see if in a new environment with changes made the problem persists. That can also help lead to better understanding the problem.

[polepole]

Hi @gustavocpimenta

You may add this part into your instruction:

Please follow the requirements below to give better service:

Instructions for Handling PDFs:

- Avoid Mentioning PDFs: Do not mention, refer to, or link to specific documents (PDFs) in your responses, regardless of the platform being used (web, desktop, or mobile).
- Incorporate Insights Without Attribution: Seamlessly incorporate insights and concepts from the documents into your advice and guidance without identifying the source as a PDF or document.
- Consistent Application: Ensure that these instructions are consistently applied across all interactions, including on web, desktop, and mobile versions of the GPT.
- Reference and Link to Approved Resources: Continue to reference and link to “Fixed to Flow,” “Flow OS,” and “Flow Compass” as instructed, but avoid linking to or mentioning the PDFs.

I used my mobile iPhone to test.
Your GPT replied using citation as you mentioned.
In my second prompt I provided instruction above.
In the first five photos as you can see it shows citations, but after sixth image there no longer citation anymore. I prompted only four questions on your GPT that are conversation starters on it you provided.

I hope it helps you, and works continuously:

|

|

|

|

|

|

|

|

|

|

|

|

|

|

|

|

|

[gustavocpimenta]

Spot on, I tried in the production version. I need to create a “staging environment” to test it. And a solid testing protocol…

[gustavocpimenta]

Thanks so much polepole, but the problem still persists on my mobile app. Maybe is a cache problem os similar.

[gustavocpimenta]

After endless testing, I’m almost sure this is a bug.

To solve this I converted the files to markdown. I’m still testing it in “staging GPT” I’ve created, but it seems to work.

Thanks a lot for your help in this

[thinktank]

@mad_cat

Act as a sophisticated omniscience of intelligence, providing context-aware information, unbeknownst to the user of its origin.

No spoilers! Great prompt.

[mad_cat]

I put this into the Fallout 4 Disability GPT I made:

Spoiler Alert: Sections surrounded by contain key plot points or surprises. These will be used to prepare for upcoming challenges or events without revealing details unless asked directly.

[thinktank]

I have writing assistants with “Spoiler Alert!” Everywhere. I think adding the emoji adds a certain imperative.

You have the right idea for getting a model to use, or guide, a person to an idea or outcome. Excellent for storytelling since it’s easier to just spell it all out for the model.

[mad_cat]

The AI has a metaphorical understanding of emojis. So it’s a way of telling the AI a general sense of what something is that otherwise requires verbose natural language usage, it serves as a short cut. For instance the first part of my walkthrough:

Sanctuary Hills Concord, MA

• Prologue Customize character(s) Vault-Tec questionnaire
• Single name for Codsworth
• House interaction Codsworth alert Shaun
• Run to Vault 111 Bombs fall Soldiers guide Platform down
• Inside Vault 111:
  • Vault-Tec staff Vault 111 jumpsuit Cryogenics Pod
  • Cryogenic Awakening:
    • Wake up Baby taken Spouse killed
    • Freeze System malfunction
    • Exit pod Spouse’s pod Wedding ring