#(Open Source) ChatGPT + Internet + Personas | (Quest🔍)

23 messages · Page 1 of 1 (latest)

tiny yoke
#

If you've been in this server for a while you might have seen Quest before. It is an open source project that implements the new ChatGPT API with internet searches. You can also choose how the assistant behaves, so no need to jailbreak it.

If you are wondering what changed since the last version. Simply put, OpenAI released the ChatGPT API, which is 10x cheaper and performs better on many tasks compared to davinci-003. This update simply implemented the new model.

At the moment I don't plan on continuing development of the app unless I see enough people interested in it. There are two low hanging improvements that could be done still: GUI for creating new personas and automatic SEO for better search results. Please let me know if you care to see these features implemented.

App: https://aichat.streamlit.app/

Streamlit
A faster way to build and share data apps

This is a web app that integrates GPT-3 with google searches.

silk saffron
#

May I be the first to have an interest with your project.

#

At least this month.

#

There should be a function to start over the chat when someone resets their API key, instead of saving the chat

tiny yoke
solid cloak
#

@tiny yoke is it using turbo?

snow arrow
#

@tiny yoke this is awesome, thank you for sharing it with the community. I have no idea why but... when running locally the "remember API key" box remains greyed and will not allow me to set

#

Update: added the api key as value= in st.text_input

tiny yoke
tiny yoke
snow arrow
#

This is awesome, just going to have to put in some time to create + manage the database locally. Thanks @tiny yoke I'll get you a coffee sometime

#

@tiny yoke is deta.sh really that bad?

tiny yoke
# snow arrow This is awesome, just going to have to put in some time to create + manage the d...

If you mean a database for the internet information and past conversations, then I guess you can take a look at the database.py script. It's mostly commented out, but if you create the implementations, it should work.

As for Deta. It was not handling many of the requests properly at the time this app was more popular. I don't know if they improved their backend, but at this point, I would just create my own custom implementation. Probably using an open source vector database like Milvus.

snow arrow
#

Yep just reading through, I only plan on running locally here at my home so this will work

tiny yoke
# snow arrow Yep just reading through, I only plan on running locally here at my home so this...

You can most likely get by using something simple like SQLite. But if you think you will be searching for thousands of messages, then it's best to use Milvus. With a tabular database you would have to load all embeddings into memory, which can be slow and resource intensive for a really big number of messages. In that case, a vector database would be a better (albeit a bit more complicated) solution.

snow arrow
#

You really did a great job with this

tiny yoke
shrewd cipherBOT
#

@tiny yoke - High-risk word(s) or topic(s) detected

Farrael'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.
snow arrow
tiny yoke
#

Btw, I made a couple of bug fixes. One associated with the estimated max tokens allowed for completion (It was being overestimated due to the new way tokens are calculated). I applied a band-aid fix where the maximum estimated request is now 3000 instead of 4000 tokens.

A second bug happened when asking duck duck go certain questions and it would return None breaking the code. This scenario is now properly handled.

tiny yoke
#

I've been investigating the LangChain Agent class for dynamic use of things like calculators and search APIs. It turns out it is very unstable when trying to integrate different personas with it. Prompt engineering is so annoying sometimes.