Anyone having issues with a GPT conversation that has problems creating word documents? E.g. Working with some structured sample data in a CSV file (500 rows) asking for some basic analysis then requesting the analysis is saved to a Word file for future reference in my project.
This has worked in the past e.g. via GPT4 and a GPT.
Thanks
Paul
Unfortunately, I’m not sure about how it creates word documents precisely. How did you achieve this before, and what is it doing now?
I’ll be honest I don’t typically ask for a docx file directly from it; I find it has a much easier time producing a .txt file instead, plus txt files are extremely easy to work with. If it were me, i’d just have it produce the doc as a txt file and copypasta it into a word doc (or whatever doc you wnat). I’ll also mention that I prefer this method because word docs have weird encodings / metadata. Word docs aren’t plaintext like txt files, there’s extra “gunk” in there. Perhaps it’s having trouble because of this? That’s my best guess, assuming this isn’t a prompt issue (which it very well could be).
2 Likes
Maybe that would have worked with code interpreter and a python module. Do you have it enabled?
Ask code interpreter to list python sandbox library modules and find ones that can create a document compatible with Word. Then add a bit of information about that ability in your system instruction.
3 Likes
You can take a look at the error message when clicking the “show work”/arrow symbol. Which may already be helpful to identify the root cause.
Adding this because yesterday I ran into issues to retrieve the list of installed modules.
And you can go back to an earlier conversation when it worked and check the code for a comparison.
1 Like
Thanks everyone for replying.
I figured that if I configure the GPT instructions to be great at producing word docs, & saving sql scripts to .sql files that might work.
I was guessing but thankfully it did.
I think (I am a complete newbie in discovering GPT power) that a lot of what we can expect from a GPT will be based on how we customise it.
So far so good.
1 Like
seems to be a guide in how much it “wants” to write.
I dare not speculate beyond that, save to say, I have tried may ways with much nuance to get it to write more.
It seems biased to write small tidbits of what you want.
I uploaded a lesson plan Word template as a sample for ChatGPT4 to reference. It has a table with 5 named columns e.g. learning activities, duration etc. Then give instructions on generating a downloadable Word document using the same table format as the template for a new lesson plan for a specified topic. The outcomes are erratic and gets weirder each time I refine my prompts:
- First time it generated a downloadable Word document but the lesson plan is simplistic (despite giving it an instructional framework and expected outcomes).
- Second time it just gives me back my template albeit in the form of a new Word document but empty inputs.
- Third time it says that it needs a lot of time to prepare the document but nothing comes out.
- Finally, it throws “its hands up” after several attempts to analyze the template and says that it is not able to produce Word document.
My question for Macha is: If I ask it to produce a .txt file instead, how do I reproduce the output in Word such that the text can be displayed in table format? If this is not possible, could an alternative be a csv file that I can open in Excel?
1 Like
Hey there and welcome to the community!
So, this is actually a perfect example of what i was describing. Word templates, especially with things like tables in them and whatnot, have a bunch of extra code mixed in we can’t see (but GPT sees), and that’s pretty much why you’d be getting that erratic behavior. Word has some weird special optimization hoo-ha which is why things like tables in a word doc are a nightmare to deal with (which isn’t specific to ChatGPT problems btw).
While it’s possible to solve your problem with a txt file, I’d actually suggest your csv pitch first for your specific use case. CSVs are easy to read for Excel (and to copy/paste into an actual excel doc). Considering csv’s are also very commonly used in data analysis (and, well, the tool is called advanced data analysis), it should have an extremely easy time making you a csv file with 5 columns. You would need to describe its logic flow and be specific about how you want to fill in the columns, but it’d nonetheless be very easy for it to manage. It’s actually easier than a txt file for that tbh, because it can automatically write a script that handles the generation (or parts of it) programmatically, which should allow it to manipulate a large-ish set of data in full.
So, try that out and see how it goes.
If you run into issues though, please start a new topic so we can focus on your specific use case and problem solve from there.
2 Likes
Thanks Macha. Sounds like a plausible route. Will give it a try and share the outcomes.
1 Like
Thanks Macha, I’ve actually had a modicum of success in saving the chat content to a word doc , however it still leaves out detail. So for now I just copy from the chat UI into the document, formatting is not so great of course as the chat style is retained.
Saving to documents is useful especially when we (data analysts) are building up valuable documentation that otherwise does not exist within the project.
1 Like
Good to hear you got something that works for you!
These models can definitely be finicky, so it is a learning curve to get more precise outputs in the way that’s desired. Also, I’m gonna assume by format you mean the genuine query - response format more broadly, not the actual format of the pasted text, which can be removed via paste options when right-clicking. And that’s not to be snarky, I did actually spend like a good 5 minutes the first time I pasted text from ChatGPT wondering why it was formatted weird, and then I discovered I had different ways I could paste text lol. Figured I’d leave that tidbit there just in case anyone else needs it lol.
1 Like
Yes I did mean the genuine query response, I admit I did not actually think to look at the paste options, I’ll look at that next time.
Definitely a learning curve, but this I find as exciting as when I first started programming in my early career (a bloody long time ago 
1 Like
What has your experience been since? Chat has been creating fantastic word docs for me from content developed in chats, up until a few months ago… Created a formatted template for job postings I’d edit and request a new posting in word, when ch it easily generated and provided a link to.
Then, a couple of months ago this functionality disappeared, along with several others, like transcribing audio and video interviews, and more.
Last nights chat got asked for audio file to transcribe then stated unable to. Investigating these losses further we tested several processes / functions it thought it could do only to learn it can not. AND, that when a function it’s been trained to do fails it has no method of reporting failures.
Learning each session is a blank slate wherein it loads user experience markers to tailor interaction styles and “recall items such is company and personal topics (I often request it commit information from a chat to memory so it is prepared in our next session.) such as business profile, writing styles, likes, dislikes etc., as well as uploaded several pieces of work to learn these styles and refine them.
Back to the loss of functions. CGPT states has no self reporting structure it is aware of. However “we” have implemented (untested) a method to remind it of these challenges and check them on initiation of a chat! Given it maintains a user file imported at the beginning of each session it has added a user preference trigger that it verify various functions it thinks it has with the goal of reporting in its creating to me, the status of same. This also triggers it to cease offering functionality which no longer works.
While it’s trained to “ learn” (through these profile markers being loaded to its blank natural state each session) communication style and preferences, there’s no function of recalling or reporting failures (that it is “aware of”. We’re now testing this config we worked through and will see.
Correlation: without noting specifics, I feel the timing of these losses of functionality align with the Microsoft acquisition of or investment in ChatGPT to integrate into its app’s. Something I’ve not yet worked with.
1 Like