# CustomGPTs - poor results querying stored PDF & .txt files

**URL:** <https://community.openai.com/t/customgpts-poor-results-querying-stored-pdf-txt-files/927446>\
**Category:** GPT builders\
**Tags:** pdf, custom-gpt, custom-gpts\
**Created:** [August 31, 2024, 3:07am UTC](https://community.openai.com/t/customgpts-poor-results-querying-stored-pdf-txt-files/927446 "2024-08-31T03:07:20Z")\
**Posts on this page:** 3\
**Page:** 1

<div class="post-metadata">

**Author:** ![johnfromberkeley](https://sea2.discourse-cdn.com/openai1/user_avatar/community.openai.com/johnfromberkeley/32/448547_2.png) [@johnfromberkeley](https://community.openai.com/u/johnfromberkeley)\
**Post date:** [August 31, 2024, 3:07am UTC](https://community.openai.com/t/customgpts-poor-results-querying-stored-pdf-txt-files/927446/1 "2024-08-31T03:07:20Z")

</div>

I have some customGPTs where I have uploaded PDFs and dot TXT files. I get extremely poor results querying those documents. They have trouble finding things, and they hallucinate like crazy with results. Is this typical?

By the way, I am talking about the ChatGPT pro GUI custom GPTs.

---

<div class="post-metadata">

**Author:** ![LikelyCandidate](https://sea2.discourse-cdn.com/openai1/user_avatar/community.openai.com/likelycandidate/32/448274_2.png) [@LikelyCandidate](https://community.openai.com/u/LikelyCandidate)\
**Post date:** [August 31, 2024, 5:00pm UTC](https://community.openai.com/t/customgpts-poor-results-querying-stored-pdf-txt-files/927446/2 "2024-08-31T17:00:54Z")

</div>

@johnfromberkeley there are two possible areas to investigate:

1. Do the text files contain adequate information to answer the prompts?  
Run the customGPT and investigate using these two prompts:  
• Categorise all the knowledge in the files provided, with a title and 30 word description.

- Using the files only, give me 20 questions the user can ask

1. Is the custom GPT answering using the files, or just ChatGPT knowledge?

- Add this to your CustomGPT prompt: ALWAYS answer using the files provided. If you cannot answer with the files provided, inform the user you don’t have the knowledge available.
- Then to see the source used for answers: For the user questions and responses above, create a a table, with columns “Question, Response, Source (Files or World knowledge)”

---

<div class="post-metadata">

**Author:** ![johnfromberkeley](https://sea2.discourse-cdn.com/openai1/user_avatar/community.openai.com/johnfromberkeley/32/448547_2.png) [@johnfromberkeley](https://community.openai.com/u/johnfromberkeley)\
**Post date:** [September 1, 2024, 10:14pm UTC](https://community.openai.com/t/customgpts-poor-results-querying-stored-pdf-txt-files/927446/3 "2024-09-01T22:14:43Z")

</div>

Thanks. Re: 1 yes.

Re: 2, thanks, I’ll try that and report back.
