How to build a ChatGPT application with multiple PDFs using Flowise

[@JoaquinTorroba]

Hey Derek (and everybody), any idea how to link this "pdf chatflows" with a website or web-app (such as bubble)? I tried making the CURL call from Bubble's API Connector plugin, and I can chat with the PDF, but the upload is from Flowise, not from Bubble. I'm looking to integrate it so that uploads can be made from the web app. Thanks!

[@derekcheung2598]

Hi Joaquin, Flowise allows you to embed right into website. I created a video for that in my channel 😀

[@wasima4463]

what if i ihave hundreds of pdfs and if the number is growing dynamically? do I need to manually wire them? if yes then its quite inefficient. if no, then how do you properly display such dense connections? that sort of nullifies the whole point.

[@AIEntusiast_]

have same question, and seems you have to do 1 by 1

[@javierfuentesmora1814]

​@@AIEntusiast_ I have the same question 😮 did you find any answer?

[@surajkhan5834]

Greate..

[@Nexus3NL]

Hi Derek thank you for your vid! how would this flow be when you also want info from external sources in combination with the pdf?

[@DerekTruong472]

Great question Ludmilla. You can add from other sources as well since there are other built in document loaders. What sources are you most interested in?

[@mikew2883]

Great video! Quick question that I cannot find an answer to anywhere. In Pinecone's documentation, they advise users to delete their index after use. While it might seem counterintuitive since an existing index could offer faster queries, the reason behind this recommendation likely relates to resource consumption. Pinecone charges for index usage, even when it remains unused or contains no data. Deleting unused indexes helps optimize resources and cost-effectiveness, especially for enterprise solutions dealing with a large number of vectors to search on. This does not seem like an enterprise solution to me if you are have to possibly embed hundreds of thousands of vectors for each query you perform, does it?

[@jd2161]

That's doesn't sound correct. You must be misunderstanding as that would defeat the entire purpose of storage

[@jessebusma5339]

How can we automatically add a pdf to vector store every time a member of my team uploads a pdf into a google drive?

[@consig1iere294]

The verctor service is giving me an error, "Request to Ollama server failer: 500 Internal Server Error". I checked, Ollama is running fine.

[@nntun03]

superb video! thanks..may I know if I will smthg for it to remember the dialogue? memory? TQs

[@sylestra]

hi there, after saving the flow I'm getting this , any idea what that could be? thanks

[@bourbe]

Hello, I am wonderng About something, when WE se a csv agent, WE don't need to use embeding, Vector data base or a memory ? I am currenly confuse

[@thewixwiz]

Hi Derek, thanks for the tutorial! I'm curious if this will upsert the PDFs to the Vector database with each call to the chatbot or only the first time? Is there some kind of caching going on? Thanks!

[@derekcheung2598]

Just the first time. There is indeed some caching.

[@musumo1908]

Great video again. Any reason you dont use the upload files by folder option? Thanks

[@derekcheung2598]

Hi Mu, If you want to add in meta data into specific pdf it's easier to add individually like we did in the example. But if no meta data needed, folder option can make it easier for many files.

[@jawadafzal6170]

​@@derekcheung2598 Hi, I'm running my Flowise in render, and I have many PDFs in one folder. When I use 'Folder with files,' it asks me to input a path, but the problem is when using it with Render hosting, it never detects my path and gives me an error. I want to add a folder because I have multiple files.

[@chinchilla989]

any idea how to prompt chatGPT in this case to get a more personalized result?

[@leonamnagel9999]

Can you explain about the Recursive Character TextSplitter Settings? Why 500 and 200?

[@JoaquinTorroba]

My understanding: when you upload the pdf you need to convert the whole text into chunks. Example: The text has 10,000 words, you can split it into 10 chunks of 1000 words each. For each chunk you need to define 2 things: 1. Size of the chunk: In the example this would be 1000. The size matters cause then this chunk goes as "context" to the LLM. I mean the question the user asks + the chunk as context so the LLM can answer correctly. BUT, you should consider the size cause there's a limit in the quantity of tokens you can send to the LLM (question tokens + chunks tokens = total tokens). 2. Overlapping: If the answer to the question needs 2 chunks that are in the same part of the text, with overlapping you can make it more efficient and it can send 1 chunk + overlapping so you don't overload the quantity of tokens (in the example above instead of sending 2 chunks of 2000 tokens, you'll be sending 1200 tokens).

[@mikehynz]

@@JoaquinTorrobaThis answer is amazing.

[@nishantkumar-lw6ce]

Can you do something with hugging face models and embeddings?

[@derekcheung2598]

Hi Nishant, yes it is possible to do this with Flowise by using the HF LLM and then specifying the model. Generally, speaking the HF models that you can point too are much weaker than OpenAI. So I don't use it.

[@jawadafzal6170]

Hi, I'm running my Flowise in render, and I have many PDFs in one folder. When I use 'Folder with files,' it asks me to input a path, but the problem is when using it with Render hosting, it never detects my path and gives me an error. I want to add a folder because I have multiple files.

[@thecolony1480]

Did you solve this? I want to do the same thing!

[@OFWCREATOR]

What if I have 20 gb worth of pdfs

[@nishantkumar-lw6ce]

The outputs of these built using flowise is very bad. Have you noticed the outputs generated here?

[@nishantkumar-lw6ce]

How to get an accurate output for a PDF for any document?

[@derekcheung2598]

Hi Nishant, getting good results on the answers are a couple of factors: the size of the chunks and the model you use. I recommend turning on the "return source document" button so you can see what the LLM sees when you ask the question. And then iterate through the chunk size as well as the K number from the vector store.

[@nishantkumar-lw6ce]

@@derekcheung2598 does flowise support the source option as well as K parameter? Usually I use K=3 while I code

[@derektruong4763]

@@nishantkumar-lw6ce and it also supports showing source doc

[@MichaelKaspy-n3r]

Great video! Do u know if there is a way to remove powered by flowise or customize it?

[@derekcheung2598]

Hi Michael, I describe how to customize the chatbot in this video kzbin.info/www/bejne/iYHFkoqaha-HocU

[@MichaelKaspy-n3r]

Yes thats the video i commented lol.. I have seen it but i dont see how to remove poweredby or change the text.. I only see the possibility to change the colour of that text. Do u have any suggestion?

[@derekcheung2598]

@@MichaelKaspy-n3r You can make the watermark disappear by setting poweredByTextColor: "#ffffff"

[@MichaelKaspy-n3r]

Thanks a lot!

[@ghulam-e-mustafapatel4894]

If we upload 10 file so it will work or not?

[@derekcheung2598]

Yes, it will

[@rhinoclark]

I'm receiving an 'Error: request failed with status code 429'. I assume it's because I'm using free versions of Pinecone and ChatGPT. Any ideas on a workaround?

[@darknessguy4221]

Use, you need to upgrade or set your OpenAI billing to get access to the API.