OmniplexAI: The Universal Computation Suite

Current Architecture of the project
Example of a Stock fund manager using the OmniplexAI as base
TiDB Dashboard
Tools available
Compute utilised by the notebook
Example 1(1)
Example 1(2)
Example 1(3)
Example 1(4)
Example 2 (1)
Example 2 (2)
Example 2 (3)
Example 2 (4)

Inspiration

The inspiration behind the project, drawing parallels from Jarvis in the Ironman series, is to create a highly advanced AI tool that goes beyond capabilities of conventional llms. The project aims to develop an AI-driven system that acts as an indispensable aid for users, assisting them in various tasks and decision-making processes.

What it does

Our project is a proof-of-concept which ultimately aims to create a highly versatile and powerful computation tool with the ultimate goal of handling a wide array of problems. At its current stage, the tool is already equipped to provide insightful answers to queries, perform sentiment analysis, and efficiently store data using TiDB. For POC, we have added a use case of offering stock trading signals.

How we built it

Our project draws inspiration from the open-source community, leveraging the power of open-source transformer models and Langchain technology to develop our computation tool. Through the utilisation of open-source transformer models and the innovative Langchain, we have laid the foundation for a cutting-edge and inclusive computation platform, promoting transparency, collaboration, and continuous improvement in our journey toward a more capable and versatile tool.

Challenges we ran into

During the development process, we encountered several challenges that required innovative solutions:

Adapting Giant AI LLM Models to Resource-Constrained Environments: One of the major hurdles was running large AI LLM models in resource-constrained environments while ensuring optimal performance. We needed to find ways to make these models work efficiently with limited resources without compromising their capabilities.
Integrating OpenAI's Langchain Library with Other Open-Source Models: Another significant challenge involved seamlessly incorporating OpenAI's Langchain Library, designed for ChatGPT, with other open-source models. We aimed to create a cohesive and adaptable ecosystem where Langchain could interact with various open-source models, enabling enhanced functionality and collaboration between different AI components.

Accomplishments that we're proud of

We take immense pride in our noteworthy achievements throughout the project:

Integrated TiDB with Langchain: Uses TiDB to store and manage large language models used in the Langchain platform. TiDB's scalability and automatic data sharding can be beneficial for handling large-scale language models. Saving natural language data with timestamp.
Efficient Utilization of AI Models: Overcoming the challenge of limited resources, we successfully managed to run two AI models, including the latest llama 2 from Meta and Google's Bert for sentiment analysis, efficiently within the constraints of under 11GB video RAM.
Universal Adaptation of Langchain: We achieved a significant milestone by extending the capabilities of OpenAI's Langchain to be compatible with any open-source LLM (Large Language Model). This achievement allows seamless collaboration and integration with various LLMs, fostering a more versatile and interconnected computation environment.
Tackling Model Hallucination: We successfully implemented measures to address the issue of model hallucination, ensuring that the AI models remain focused on the problem at hand by prompt engineering. By mitigating this challenge, we enhance the reliability and accuracy of our computation tool, ensuring trustworthy results for users.

Tools integrated:

Google Search Google Search is a powerful web search engine that allows users to find information on the internet. It is widely known for its comprehensive index and accurate search results. In this integration, we can leverage the Google Search API to programmatically retrieve search results for specific queries.
DuckDuckGo Search DuckDuckGo is another web search engine that focuses on user privacy and not tracking user data. It provides search results from various sources while respecting user anonymity. By integrating DuckDuckGo Search, users have an alternative search option that prioritises privacy.
Google BERT Sentiment Analysis BERT (Bidirectional Encoder Representations from Transformers) is a state-of-the-art natural language processing model developed by Google. It is capable of understanding the context and nuances of text, making it ideal for sentiment analysis tasks. In this integration, we can use Google's BERT model to perform sentiment analysis on textual data, determining whether the sentiment is positive, negative, or neutral.
TiDB Data Store TiDB is an open-source distributed SQL database that enables scalable and highly available data storage and retrieval. It is designed to handle online transaction processing (OLTP) workloads and provides horizontal scalability by adding more nodes to the cluster. In this integration, we can use TiDB as a reliable and distributed data store to save and manage various types of information.

Examples:

2 examples are attached in the pictures of the project

Example 1 shows how you can give connecting tasks to the controller.
Example 2 shows that the controller can perform more than 2 tasks in a given query. performance can be improved by prompt engineering.

Limitations:

We are currently limited by the compute which is available to us.
Larger models will perform better.

What we learned

We learned valuable lessons and gained insights that have enriched our knowledge and expertise in several areas:

Resource Optimisation
Integration of Open-Source Tools
Connecting TiDB to AI tools

What's next for OmniplexAI: The Universal Computation Suite

Our objective is to make this project open-source, enabling the community to contribute and add more tools to the platform. Additionally, we aim to empower users to extract AI software stacks capable of performing specific tasks by breaking them down into smaller tasks and utilising various fine-tuned models or even using the same model with different personas.

Our next objective is to integrate this AI model with AWS services, particularly Lambda, to transform it into an event-driven project. By doing so, we aim to create a streamlined and efficient system that automatically processes user inputs for sentiment analysis. Additionally, we plan to develop a user-friendly dashboard that will enhance the overall user experience, providing a simple and intuitive interface for interacting with the sentiment analysis tool. Through this integration with AWS and the dashboard development, we anticipate optimising the accessibility and usability of our solution.