Thanks. I appreciate the comment!

Thanks for watching! Let me know if there are any other topics you would like to see.

You’re welcome! Let me know if there is any specific content that you would be interested in seeing.

I’m glad you enjoyed it. Thanks for the comment.

In the model give a system instruction like, (system instruction= "Give response only from the file uploaded and not from outside the pdf or csv file provided" ) . This will limit it to answer only from the knowledge base given, feel free to customise your system instructions.

There are 2 approaches you can take. One is to just put all your data inside the prompt. This is possible nowadays even with big datasets as models now have huge context windows (up to 2 million tokens for Gemini 1.5 Pro). However, adding too much data to the prompt can be costly and even result in lower quality answers. The other approach is RAG (Retrieval Augmented Generation) which extracts smaller chunks of your dataset that are similar to your input query / question and then uses that data to produce an answer. This is much more efficient and when done properly may even result in better outputs. With either method, you would include in the prompt an instruction to only consider the data you provided when generating the response. This doesn’t work 100% of the time but overall it’s pretty reliable. I put out a video a little while back that demonstrated RAG for a chatbot and how to restrict the model to the data you provide. Check out the video here: kzfaq.info/get/bejne/mdadZZuZqtyrcXU.htmlsi=qAX6wM0ytI_VSu7v

Yes, definitely! Gemini 1.5 is multimodal and can take images, audio and video in the prompt with text. I am working on a video on Gemini for vision / images right now and will let you know when it’s posted.

@@aarondunn-zt7ev okay, ty

I just uploaded my video on Gemini for vision. Check it out here: kzfaq.info/get/bejne/jsl9lrFzv8fbcYk.htmlsi=6p93NAlB6o9DePnz

Thanks. Happy to hear it was useful.

There are 2 approaches you can take. One is to just put all your data inside the prompt. This is possible nowadays even with big datasets as models now have huge context windows (up to 2 million tokens for Gemini 1.5 Pro). However, adding too much data to the prompt can be costly and even result in lower quality answers. The other approach is RAG (Retrieval Augmented Generation) which extracts smaller chunks of your dataset that are similar to your input query / question and then uses that data to produce an answer. This is much more efficient and when done properly may even result in better outputs. With either method, you would include in the prompt an instruction to only consider the data you provided when generating the response. This doesn’t work 100% of the time but overall it’s pretty reliable. I put out a video a little while back that demonstrated RAG for a chatbot and how to restrict the model to the data you provide. Check out the video here: kzfaq.info/get/bejne/mdadZZuZqtyrcXU.htmlsi=qAX6wM0ytI_VSu7v

I don’t have any experience creating Facebook apps, but I can look to do some more research on it. From what I’ve discovered so far, to add a Gemini-powered chatbot to your Facebook page, you'll need to integrate it using Facebook Messenger's API. This involves setting up a Facebook Developer account, creating a Facebook app, and configuring a webhook to handle messages. You'd then connect the Gemini API to process and respond to these messages. It would be an interesting project and video!

To run your chatbot on Google Cloud Run, you'll need to containerize your application using Docker, then deploy it. This involves creating a Dockerfile to package your code and dependencies, pushing the image to Google Container Registry, and deploying it via the Google Cloud Console. You can follow Google's Cloud Run quickstart guide (cloud.google.com/run/docs/quickstarts ) for detailed steps. If there's interest, I could create a detailed tutorial video on this process in the future.

@@aarondunn-zt7ev that would be great if you could make such a tutorial. I have tried to put it on cloud run but my understanding is lacking. I keep getting a "service Unavailable" message on my run url. Im also not sure if I need to get a secret key etc. (my program runs fine from the console, but fails from the cloud run url). Anyway, great tutorials and I have subscribed :)

Thanks for the comment. I appreciate it! As of this moment there is no SQL sanitation that would filter out malicious queries. This is a very good point, and something could be explored. One of the factors that would protect the database from injections is that it’s the LLM which creates the query based on the prompt and there are constraints in the prompt which restrict what can go into the query. However, I’m sure that some very creative prompt injection strategies could have the LLM produce unexpected and undesirable behaviour. Its something to experiment with and research further.

@@aarondunn-zt7ev at the very least I suppose limiting the access of the user running the LLM's queries would prevent any little Bobby tables incidents 😅

@clint5724 I don’t have the app available publicly but this is definitely something I can do if there is interest. I will look to get it online potentially over the next few weeks. Stay tuned!