FreeStream

Try out different AI tools.

TLDR:

• A Streamlit multi-page application
• AI chatbots for different scenarios:
  • RAGbot - Have an AI generate text based on what you upload
    • Your uploaded files are deleted upon exit
  • Chatbot - Work with a general purpose AI assistant
• All conversation content is traced via LangSmith for developer evaluation
• No sign-ups required

Table of Contents

• Quickstart
  • Installation
• Description
  • Vocabulary
  • Current Functionality
• Functional Requirements
• Non-Functional Requirements
• Roadmap
  • Thinking Out Loud
  • Future Functionality Plans
• License
• LLM Providers' Privacy Policies

Quickstart

This app is hosted via Streamlit Community Cloud, here

Installation

This project uses poetry for dependency management because that's what Streamlit Community Cloud uses to deploy the project.

Install it with:

pip install -U pip && pip install -U poetry

Then, install the project's dependencies in a virtual environment using poetry.

Run:

poetry install

You will need to set all required secrets, which require their own respective accounts. Make a copy of "template.secrets.toml" and rename it to "secrets.toml" in the root of the project. Fill out each field in the file.

Need API Keys?

API Platform	Link
Claude	https://console.anthropic.com/
Google	https://aistudio.google.com/app/apikey
Langchain	https://smith.langchain.com/
OpenAI	https://platform.openai.com/api-keys

You can then start the development server with hot reloading by running:

poetry run streamlit run ./freestream/🏡_Home.py

---

Description

I originally created this project as a chatbot for law and medical professionals, but I quickly realized a more flexible system would benefit everyone.

Important Vocabulary

In order of importance, specific to our learning.

Vocab	Definition
Large Language Model	A model that can generate text.
RAG	Retrieval Augmented Generation
C-RAG	Corrective-Retrieval Augmented Generation
Self-RAG	Self-reflective Retrieval Augmented Generation

What can FreeStream do for me, currently?

FreeStream functions as a multi-page Streamlit web application, which means it's running on typical web technologies like JavaScript but its actually built using Python.

There are "bot" pages where you can interact with an LLM of your choosing, for example, GPT-4 or Claude Opus. These pages have vastly different prompt engineering techniques being applied, so you may specifically find the prompts in the code interesting even if you aren't a programmer. The first bot page was "RAGbot," which allows you to upload a file (or files) and ask it questions.

For now, the only non-"bot" page is Real-ESRGAN, an image upscaler trained on "pure synthetic data" that can upscale to arbitrary ratios in GPU-supported environments.

Functional Requirements

The application MUST...

1. Provide a user interface for chatting with the latest large language models.
2. Leverage advanced prompt engineering techniques.
3. Provide a range of chatbot pages, differentiated by their prompt engineering.
4. Let the user "drop-in" their choice of LLM at any time during a conversation.
5. Allow users to perform image upscaling (PDF, JPEG, PNG) without limits.

Non-Functional Requirements

The application SHOULD...

1. Aim for 24/7 availability.
2. Prioritize ease of navigation
3. Feature a visually appealing seamless interface.
4. Empower non-technical users reactive AI.
5. Let users generate tasks based on their speech

Roadmap

Thinking Out Loud

The next bot pages I'd like to make are an AgentExecutor and an advanced state machine using LangGraph. For the AgentExecutor, users could have a general purpose assistant that's more robust than "Chatbot". And for the state machine, I could implement corrective RAG for the first time, or maybe do something like RAPTOR or ColBERT.

Future Functionality Plans

• Create an RAG chatbot
• Create a general purpose AI chatbot
• Add Gemini-Pro to the model list
• Add Anthropic's Claude 3 model family
• Add AI decision making
  • Implement Corrective-RAG OR Reflective-RAG
• Turn into a Multi-Page Application (MPA)
  • (Homepage) Add a welcome screen with...
    • a description of the project
    • privacy policy
  • (Page) Migrate RAG SPA code
    • Add "Temperature" slider
  • (Page) Add "Image Upscaler"
    • Multi-file upload
    • File type detection
  • (Page) Real-ESRGAN Image Upscaler
    • Review HuggingFace Spaces's as a potential solution
  • (Page) Add a "Task Transcriber"
    • Transcribes microphone input
    • Use LLM to identify each and every task while grouping some to avoid redundance
    • Generates text within a predefined task template, for each task identified

---

License

LLM Providers' Privacy Policies

• OpenAI Privacy Policy
• Google
• Anthropic
• Streamlit

Additional References

Reference	Description
ColBERT	Efficient BERT-Based Document Search
RAPTOR	Recursive Abstractive Processing for Tree-Organized Retrieval