Overview: The AI Toolbox Plugin is the most complete and capable plugin available for creating fully customizable AI Bubble application. It allows for full customization of model settings and has a variety of models provided. At this time, it has models from OpenAI, Anthropic AI, Baidu, Google Generative AI, Google Vertex AI, Mistral AI, Together AI, and Yandex GPT.

*Important Note on Model API Keys: This plugin includes a custom client-side implementation of LangChain that requires the api keys for the available LLM models to be public. This allows the keys to be available for client-side actions and removes reliance on server-side workflows, which can diminish performance. For production apps we recommend passing api keys from the database (with proper privacy rules set) and not directly adding them in the plugin settings.

The AI Toolbox Plugin provides the ability to remember chat history and to modify model's thinking with premade and custom prompts. The prompts are used to further customize the AI with chains and agents to add additional thought patterns and tools the model can utilize.

Editor Link: https://bubble.io/page?name=index&id=silva-ai-toolbox&tab=tabs-1

How to Use:
Add the element to your page. Customize the settings on the element to select the model and change the settings. Add the action, Generate Response, to get a response from the model. Responses will be returned as part of an exposed state from the element. To know when a response is finished an event trigger will occur called Finished Response.

Components:
Chatbot Element - adds the chatbot models to the page. Contains all the models and settings for customization.
   - Actions:
     Generate Response - Generates a response from the model using the settings from the element and loaded models.
     Load Chat History From Bubble Database - Loads the chat history and inserts it into the data for the model to access when generating a response.
     Load Multi Prompt Chain - Turns the model into an AI able to be guided from multiple prompts and decide which is best to use to generate a response based on the context of the conversation. It follows a preplanned process and will take the same route for thinking every time.
    Load Agent - Turns the model into an agent that has access to tools like web browsers, search engines, and calculators to extend the models abilities. It makes decisions on the order to use these tools and is free thinking and may use a different order every time.
    Load Text Embedding Model - Loads a model that can convert documents and web pages into indexable text that the model can access as part of its memory. Useful as part of the web browser tool for agents.

Future Updates:
   More Models - Addition of several LLM models and a few more chatbots. AI21, Aleph Alpha, Replicate, Watson X, Writer, Cohere, Tongyi Quianwen (Qwen).

   Additional Chains - Document Retrieval Chains to load documents for a model to use. Structured Output Runnable to be an intermediary chain for altering the structure of outputs before sending them as part of another step.

   Custom Output Parsers - Custom Structured Outputs to alter the general response of each model. List Parsers to generate list responses. JSON Output Parser to take responses and format them as a JSON object for more structured data. HTTP Response Output to send responses as an HTTP Response as part of an API. XML Output Parser to convert XML responses to JSON objects.

   Extra Agent Tools - Access to Gmail, Google Calendar, Discord, Wikipedia, Wolfram Alpha, and Python.

   Vector Stores - Better Data management for chat history and private data provided through Chroma, Pinecone, Supabase, Vercel Postgres, and Redis.

   Retrieval Augmented Generation (RAG) - Full LLM knowledge augmentation with additional data. Allowing easy use of private data and updated public data.