ALP - local, open-source, knowledge-grounded conversational app

[fleet yacht]

https://github.com/rpast/ALP

ALP is an open-source, knowledge-grounded conversational AI system designed to generate responses grounded in relevant knowledge from external sources. ALP is currently in development, but it can be used locally on users' machines. The app reads chosen PDF file, has unlimited conversational memory and the ability to export conversation and source embeddings in JSON format.

ALP is designed to enhance the accuracy of responses of GPT-3.5 model related to a specific PDF document by using a retrieval augmentation technique. This approach ensures that the most relevant context is always provided to the model with user's question. ALP was created to help me and my friends manage the overwhelming knowledge base of research papers, books and notes, making it easier to access crucial information without having to read through everything. Maybe some of you will find it useful.

Models utilised:

1. text-embedding-ada-002
2. gpt-3.5-turbo

Features:

1. Conversational research assistant: Interact with and get information from loaded PDF file.
2. Unlimited conversational memory: Retain information from previous conversations for context-aware responses.
3. Support for long documents: You can upload books. The only thing that limits you is your API limit and OpenAI capacity.
4. JSON export: Export conversation and source embeddings as JSON format.
5. Retrieval augmentation: Utilise retrieval augmentation techniques for improved accuracy.
6. Local deployment: Spin up ALP locally on your machine for privacy and convenience (you just need to have Python installed on your PC).

demo:
https://github.com/rpast/ALP/raw/master/static/alp_demo.gif?raw=true

[echo axle]

would there be an attempt at internet access and token management?

[fleet yacht]

Im thinking about that but it is not on my priority list. The main intention was to run the tool locally and allow it to be hackable.

[fleet yacht]

ALP got an update. Now it allows to continue conversations that are automatically stored in a local database.

https://github.com/rpast/ALP/raw/master/static/alp_demo.gif?raw=true

ALP - local, open-source, knowledge-grounded conversational app

[dusk warren]

I couldn't make it run in Windows

[fleet yacht]

Hey, thanks for the info. I pushed bugfix to the repo. Now when you pull changes or git-clone again you should be able to run the app on Windows. I updated requirements.txt content and fixed dir structure error that prevented the program to spin up the database under static/data/dbs. Let me know if this solves your problem.

[dusk warren]

Thank you, will try, do you have plans on integrating llamaindex?

Maybe I could help

I'm been looking for a open source personal knowledge base, and maybe this is the app

https://gpt-index.readthedocs.io/en/latest/index.html

https://llamahub.ai/ They have a LOT of connectors

[fleet yacht]

I thought of adding more connectors with langchain. I havent heard of llamaindex before. How does it differ from langchain?

[dusk warren]

https://gpt-index.readthedocs.io/en/latest/how_to/integrations/using_with_langchain.html

[fleet yacht]

Still trying to figure out how this fits the langchain ecosystem. From what I see llamaindex offers data connectors and index management. But langchain offers this as well. What am I missing?

[dusk warren]

Oh I didn't know that, maybe I'm wrong

I will try your project when I have my PC again

[fleet yacht]

I mean, this doesn't necessarily mean that llamaindex isnt worth giving it a shot. I'm just not familiar with it. Anyway, I will play with it this week. Thanks for recommendation.

UI, UX upgrade:
https://github.com/rpast/ALP/raw/master/static/alp_demo.gif?raw=true

[sleek solstice]

Given that you still need to supply part of your text to gpt, there is potentially leaking concern if I am working with sensitive/confidential data. Is it fair to say so? @fleet yacht

[fleet yacht]

Yes, I wouldn't use it for sensitive/confidential data. In the future I might add support for https so it will be harder to intercept signal.

[sleek solstice]

I am not concerned about http/HTTPS. Local host is not even on network adapter…it's the small portion of text sending to gpt as part of the prompt would get me in big trouble lol

But hey, it's a good project for everything else!

[fleet yacht]

Yeah, the context is sent to gpt as a natural language snippet. I'd love to give it a support for locally stored llms like alpaca. Im experimenting with it rn. This would be ultimate safety fuse for use cases like yours. Otherwise I cannot see any other way to do that unless Open AI implement encryption feature.

[cobalt slate]

any updates on that? I would be interested in running it locally with vicuna

[fleet yacht]

hey, thanks for being interested in using ALP. Im currently revising the data model for the app. How do you envision an interface for such functionality? I mean, the support for other llms.

[cobalt slate]

Not sure, maybe a picker like the drop-down menu you see elsewhere or just a different repo for each. Doesn't really matter ad long as it works.

[fleet yacht]

new major update:

• new data model: collections<->sessions decouple
• sentence transformers responsible for embedding