June 2023 Summaries

Activeloop, in collaboration with Towards AI, has launched three free skill certification courses on LangChain, Vector Databases, and Foundational Models as part of the Gen AI 360 initiative. The program aims to educate and upskill engineers in the rapidly growing field of AI. Deep Lake, a vector database for all AI data developed by Activeloop, is designed to combine the best of both data lakes and vector databases, enabling companies to fine-tune their Large Language Models on complex data beyond embeddings. LangChain seamlessly connects Deep Lake datasets with Foundational Models in diverse use cases. The certification program will foster the adoption of cutting-edge AI technology within any engineering organization across industries.

FableForge is an open-source app that generates children's picture books from a single prompt using AI technology. It combines GPT-3.5/4 for text generation, OpenAI's function calling feature to generate Stable Diffusion prompts, and Deep Lake for storing generated images and text pairs. The app uses LangChain to interact with OpenAI's large language models and has a Streamlit UI for users to input prompts and download their generated books. FableForge demonstrates how new AI tools can be leveraged to overcome hurdles in content creation, and Deep Lake provides valuable data storage and analysis capabilities for finetuning generative AI models.

FableForge is an open-source application designed to generate children's picture books from a single prompt, utilizing advanced AI technologies such as OpenAI, LangChain, Stable Diffusion, and Deep Lake. The process begins with GPT-3.5 or GPT-4 crafting a short story, which is then converted into visual prompts for Stable Diffusion using OpenAI's function calling feature. These prompts generate corresponding images via Replicate, and the text-image pairs are stored in Deep Lake's vector database, facilitating easy storage and visualization of multimodal data. The app's development faced challenges, particularly with generating relevant prompts for Stable Diffusion using GPT-4, but function calling provided an effective solution by converting natural language input into structured data. FableForge's architecture includes text and image generation, combining them into a book format, and storing the data in Deep Lake, which supports further analysis and finetuning of generative AI models. A Streamlit-based UI allows users to generate books from their prompts and save the results, highlighting FableForge's potential in revolutionizing personalized children's literature creation.

SalesCopilot is an AI-powered sales call assistant that transcribes audio in real-time, connects users with chatbots, and provides relevant assistance within seconds. It uses GPT-3.5 or GPT-4 to detect potential objections from customers and offer recommendations on how to handle them. The initial solution of relying solely on the LLM had some flaws, such as inconsistent response quality and inefficiency. To improve this, a custom knowledge base combined with a vector database was integrated into SalesCopilot. This allowed for efficient grounding of the LLM's responses and improved retrieval performance. The final solution involved intelligent splitting of the text based on its structure to maintain semantic boundaries between each chunk. Deep Lake was used as a vector database, providing persistent storage and enabling quick and easy retrieval of relevant information.

SalesCopilot is a sales call assistant that utilizes real-time audio transcription and integrates with chatbots powered by GPT-3.5 or GPT-4 to assist salespeople with customer objections by providing relevant recommendations. Initially, relying solely on the language models (LLMs) like GPT-4 was found to be slow and inconsistent without a custom knowledge base, which led to exploring more efficient solutions. The integration of SalesCopilot with Deep Lake, a vector database, enhances the system's ability to retrieve pertinent information quickly by embedding a custom knowledge base of common customer objections and their responses. This setup involves strategically splitting the knowledge base into coherent, semantic chunks to improve retrieval performance and ensure high-quality LLM responses. The use of Deep Lake provides persistent storage, allowing the system to maintain an efficient and adaptable database for generating informed responses tailored to specific sales scenarios. This hybrid approach offers a robust tool that combines the intelligence of LLMs with the precision of vector databases, making it a powerful solution for addressing customer objections in sales interactions.

The blog post by Akash Sharma and Sinan Ozdemir explores Vellum's Playground, a solution for finding the right prompt/model mix for one's use case. They compare four leading LLMs from three top AI companies - OpenAI’s GPT-3.5 and GPT-4, Anthropic’s Claude, and Cohere’s Command series of models. The authors walk through four examples: Text Classification (detecting offensive language), Creative Content Generation with rules/personas, Question Answering and Logical Reasoning, and Code Generation. They consider three main metrics for performance/quality - Accuracy, Semantic Text Similarity, and Robustness. The goal is not to declare any of these models a “winner” but rather to help users think about judging the quality and performance of models in a more structured way using Vellum – a developer platform for building production LLM apps.