ChatPDF.com

[blazing egret]

I created ChatPDF - the easy and fast way to chat with any PDF using the new ChatGPT API, free and without sign-in. Check it out at https://ChatPDF.com and let me know what you think!

It's also live an Product Hunt: https://www.producthunt.com/posts/chatpdf
and Reddit: https://www.reddit.com/r/OpenAI/comments/11fx2cb/chatpdfcom_the_easy_and_fast_way_to_chat_with_any/)

Super excited!

[swift jolt]

Awesome! 👏 How did you achieve it, using Langchain or GPT-Index for document embedding?

[upper niche]

yoo, this is useful, thanks!

btw, how do you keep it for free? I mean, the API is paid right?

[silver tiger]

Thank you so much @blazing egret . do you plan to add paid plans? I find it very useful, gave you a vote on producthunt! keep up the work and updates here please

[weary crescent]

This is so cool

[ivory raft]

I love this! A useful improvement maybe to be able to select output language before the conversation starts. I'm having a bit of a struggle with non-english pdf's due to the first message being in the pdf's language

(also, kinda off topic, but it could also be neat to be able to upload two pdf's to talk about the differences. Like an old technique vs a new technique, or two contracts)

[silver tiger]

i think for now you could just mix two pdfs into one and you can give instructions to compare the texts inside the uploaded file.

[blazing egret]

Thanks man! It's written in TypeScript and I'm not using Langchain or GPT-index, just using an npm package to extract the content of the PDF and then create text chunks myself. It's using the OpenAI embeddings endpoint to compute vector similarity of the user question to the chunks

[blazing egret]

The API has gone so cheap since yesterday, let's see if I ever need to introduce pricing 😉

[blazing egret]

Great feedback! Will think about languages. Currently it's prompted to try to use the language of the PDF, but that might fail and the user might actually want to use a different language!

[upper niche]

okay nice! btw, this is so useful that if you ever need to have a paid to it (or in some case, perhaps use our own API key), I'll be gladly to do so

[ivory raft]

yeah I just put in an English pdf even, but it started speaking Spanish, which is strange because the pdf doesn't have anything spanish in it, the first headline is "Senior World Rankings" so maybe it thinks it's Señor (?) xD

[jaunty current]

This tbh. A document produced in my own country, even if it's in English, would return local language as the language of the chat, so I instructed the chat to use English, then it apologized and told me what help the chat would do, without translating the recommended questions to English

[jaunty current]

Nooooo need

But an own API ID would be better

[ivory raft]

yeah the first response from the bot is such a good response, but if I ask it "can you please say that again in english?" it doesn't translate the response it just writes a new one that's much worse for some reason

[solemn wharf]

Open source?

[fast badger]

It's awesome, any plan on making it open source? I want to try on a 3k pages document

upvoted 🙂

[nocturne flicker]

Who has access to the data on PDFs posted there?

@blazing egret

[neon wave]

Hi all. I provided pdf a 372-page paper (Italian language construction code, 10 Mb). Few good answers many wrong ones. However, the idea is good. Thank you for your work @blazing egret

[neon wave]

I have tried a lighter document (also in Italian) and it is much better 👍

[ivory raft]

I really like the idea. I tested with a small audio transcribe (studies), works well, all answers correct.

[ivory raft]

Personally I would not rely on question and answer format for this, you might want to avoid something too sensitive, as even in best case scenario it still goes to OpenAI after the website and I do believe they have somewhat free (or at least relatively attainable-) access to it from their hundreds of employees~

[nocturne flicker]

And that’s always going to limit it’s usefulness to me. It’s this dense stuff not documented elsewhere that I need help scanning through in the first place.

[dusk anvil]

Can it read textbooks?

[blazing egret]

Thanks for the feedback! Intriguing idea with the diff. Could you give a specific example? E.g. do you mean two contracts which almost have the same content, only some differences in a few paragraphs? And what kind of techniques have you been thinking of?

[blazing egret]

It's saved for 7 days on Google Cloud Storage and I'm the only one who has access to that. The paragraphs of the PDF are also sent to the OpenAI endpoint

[blazing egret]

Yeah sure, as long as they are in PDF format!

[zealous hemlock]

Good @blazing egret

[dusk anvil]

Nice!!

[ivory raft]

perhaps with similar content, such as a new contract from the same company (which is what made me think of this in the first place), or perhaps to compare two offers from different companies.
for techniques it could be something like a new approach or a new theory on a scientific subject, like for example comparing the old backpropagation algorithm to the new forward-forward for ML training, or two different medical intervention procedures like LASIK vs IOL for eyesight
or two competing papers like different theories on the same subject like ekpyrotic universe theory versus plasma cosmology or something like that

maybe even to compare original to a translated copy to check for discrepancies

[fresh lance]

@blazing egret please make the code available on Gitlab or Github for everyone. Your work is excellent and everybody should have this on their computer, also maybe other developers can contribute and make it even more interesting.

[latent cairn]

Can it be done for large number of related legal documents?

[ivory raft]

personally I support this being closed source 😛 one of the most valuable skills right now is prompt engineering, so if one of us makes it in this space imho it's all good, I also feel lucky to have all of this stuff for free so I don't nag openai for their closed source neither

but ofc, if mat is comfortable I might end up contributing to a repo, so, either way

[upper niche]

To be honest, I really hoped that there is an offline version available. I'd pay some bucks for it.

[fresh parrot]

That’s an unlikely scenario

you should host it on your own physical server

Actually you can build it on azure or no top of openai api

On* top

[ivory raft]

<Embedding text failed: Request failed with status code 401 > any idea what's up with this? ^^

[blazing egret]

Sorry, I accidentally disabled the wrong API key - it's fixed now!

[blazing egret]

Super interesting but seems like it would need big-picture understanding of the papers and an in-depth understanding of subject matter. Not sure the models are there yet, but I'm sure they will be soon!

[ivory raft]

Some of these use cases might be more feasible, two similar contracts from the same company, or comparing original to a translated paper~
Have you done any test? I feel like it wouldn't be too much to expect a question like "how does each paper theorise the creation of matter?" when comparing two cosmological theories for example~

[little maple]

@ivory raft

Eslof's message blocked by AutoMod

We are committed to fostering a positive atmosphere and have implemented additional filters to block any inappropriate or disruptive language and topics.

Reminder: Refrain from bypassing the AutoMod filter as per server Rule #10.

[ivory raft]

if I have some time I'm gonna see if I can try out what ((that dude earlier whose name is content blocked by the server)) mentioned e.g. combining two pdf's

[night dune]

hi. great application here. I'm developing my discord bot to have a similar feature. Would mind sharing the npm package being used for the pdf content? If my idea is correct, the process involves extracting the texts from the pdf, converting them to embeddings and storing them to database or locally.

[fading turret]

Vert Interesting and quick ! Did you have to put any safeguards/ringfences on to prevent the chatbot from spilling infos outside of what's in the PDF ?

[ivory raft]

I was able to put in a pdf going over recent findings from JWST, and got it to narrow down which alternative theory to the current big bang most aligns with the findings and I got a pretty decent explanation for why a cyclic model of the universe might cause a type of build up of mass over cycles which might lead to these abnormally large early formations~ but I had to argue with it a bit to pull comparisons from its training data instead of the PDF.
Pretty cool. I honestly think that with two papers at hand it could do well. Currently been trying to find a good example where both fit into the current limit as a single paper with some kind of divider I could prompt it to work with for testing~

[nocturne flicker]

Ok, so I think this app is amazing and good job building it. There are some large docs I want to query (200+ pages) there stuff I know I can’t share just because open AI, but there are others that are semi public anyways and this large size. However you’re asking for my API key to process the large docs. So my question is, how do I trust you? I’d like to.

[ivory raft]

Another suggestion: Download conversation button.
Say I had a really fruitful discussion about a paper and want to share the conversation with someone, at the moment I'm doing right click + save as (Web page, Single File) mht/mhtml, but a more clean export (height of the chat box gets locked with aforementioned method) or as an image or pdf or md of the conversation would be cool.

[ivory raft]

Great work here!

[ivory raft]

Maybe a 'share conversation' in either 'readonly' or... maybe even 'interactive' so that you can have colleagues or friends maybe even continue in your thread and ask their own follow up questions maybe related to the discussion itself.

[junior epoch]

Open sourced a similar one with a bad UI but better response. https://github.com/bhaskatripathi/pdfGPT

[upper niche]

Hello! Is there any way for us to insert API keys as an option? I found that there is a change of system?

[wraith horizon]

I saw this option has gone too 😦

[worthy radish]

Following

[fresh lance]

yeah, humanity takes a back seat!

[cursive stone]

there is a version of GPT that has 32000 tokens of memory. If i recall correctly 750 words is about 1000 tokens so that's 24000 words or roughly 50 pages of memory that it can work with. I imagine that it would be able to give a rather good summary of most non-review papers but it would be rather expensive when compared to the current methods which consist of summarizing chunks of text.

The issue with the current method is that in the process of summarizing chunks, some context might be lost and summaries of summaries can become opaque (from my experience). There is a mild fix to this but it is not perfect and it increases computation time and cost although it might still be cheaper than a 32000 token model (i did not compare the costs)