February 2023 Summaries

LangChain has introduced TypeScript support, enabling the recreation of prompts, chains, and agents in TypeScript with the same serializable format used in the Python version, facilitating artifact sharing between languages. This development addresses the growing interest in using JavaScript for language models, especially after the mainstream success of ChatGPT, and mirrors the Python package's abstractions, such as prompts, LLMs, and vector stores, though with slightly reduced functionality. The initiative aims to ensure that serialized prompts, chains, and agents are usable across both languages, promoting widespread community adoption and allowing prototyping in one language with the flexibility to switch to the other. While some ML-centric functionalities have limited TypeScript support, there is an intention to expand these over time, and the packages may eventually diverge to cater to their respective audiences, with the Python version possibly focusing more on research and data-centric features and the TypeScript version on web development. The community, including contributors like Sean Sullivan and Nuno Campos, played a significant role in this development, and further feedback and contributions are encouraged as the package evolves.

LangChain has announced streaming support aimed at enhancing the user experience of LLM applications by addressing latency issues through token-by-token output rather than waiting for the entire response. This improvement is particularly significant for chat applications, as it provides users with immediate feedback, similar to how ChatGPT functions. The updated chat-langchain repository now includes streaming and asynchronous execution, serving as a template for developers to build advanced chat and Q&A applications. The application utilizes FastAPI for the backend and Jinja templates for the UI and is designed to be easily deployable for multiple users. Streaming is supported for both synchronous and asynchronous executions, and callback handlers like StreamingLLMCallbackHandler and QuestionGenCallbackHandler facilitate real-time interaction through websockets. The project is open source, encouraging community input, and aims to expand streaming support to other LLMs and provide more examples for callback handlers.

A recent week-long competition aimed at developing chatbots that interact with personal data witnessed the creation of numerous document loaders and end-to-end chatbot examples. Participants developed over 25 document loaders and showcased diverse projects, each hosted on GitHub, which range from bots that analyze personal tweets and blogs to those that enable conversations with Roam Research graphs, YouTube videos, or investment-related websites. Notable entries include TokBot, which uses GPT-3 to provide insights on the TokCast podcast; Chat-Your-Data Self Hosted, which facilitates a ChatGPT-like experience over custom documents using LangChain; and various others that utilize platforms like Telegram and recipes sites. Although some submissions were excluded from the competition for not being open-source or missing the deadline, they were still recognized for their innovation. The competition will determine the winner based on the number of GitHub stars each project receives by the end of the week.

LangChain has announced its integration with Chroma, marking a significant step toward creating a modern AI stack. LangChain, a toolkit designed for developing AI-native applications, has been widely adopted for building LLM-powered applications such as chatbots and question-answering services. A critical component of these applications is the use of embeddings and vector stores, with LangChain previously relying on FAISS, which posed installation challenges. Chroma emerges as a purpose-built vector store and embeddings database, designed to be lightweight, easy to use, and suitable for local development environments, addressing the limitations of existing vector databases. The integration allows developers to prototype AI applications efficiently, leveraging LangChain’s framework and Chroma’s seamless local vector store, making Chroma an optimal choice for developers working with LangChain. The announcement emphasizes the ongoing evolution and exploration in AI, highlighting the transformative potential of these technologies in reshaping software development and capabilities.

LangChain is introducing asynchronous support by utilizing the asyncio library, which enables non-blocking I/O operations through the use of coroutines and an event loop. Asynchronous functionality is particularly advantageous for LangChain applications, which are often I/O and network bound, such as when calling LLM APIs or interacting with data stores. By allowing concurrent execution, asyncio enables multiple agents or tasks to progress simultaneously, enhancing efficiency and integration with frameworks that support asyncio, like FastAPI. Initial async support has been implemented for various LangChain components, including LLMs, chains, and agents, with further developments planned to expand async capabilities across more tools and improve functionality, such as running multiple tools concurrently and supporting callback handlers.

ChatGPT's limited knowledge scope to pre-2021 data has prompted the development of LangChain, a tool designed to help users develop language model applications over their own data sources. LangChain aims to simplify the process by providing examples and tutorials for integrating various data sources like Notion, Obsidian, and webpages. In a bid to gather diverse examples, LangChain launched a "Chat-Your-Data" Challenge, encouraging participants to create their own ChatGPT models using personal data sources. Participants are required to customize data sources and prompts using a GitHub repository as a starting point, with optional deployment on platforms like Hugging Face. The challenge not only seeks to expand the repository of examples but also incentivizes participation by offering a limited edition LangChain t-shirt to the winner, determined by the most-starred GitHub repo by February 19th.

LangChain's integration with Unstructured.io enhances its capability to combine Large Language Models with user-specific text data by leveraging Unstructured's open-source Python package. This package supports a wide array of file extensions, such as .txt, .docx, .pptx, .jpg, .png, .eml, .html, and .pdf, enabling the transformation of various data formats into machine-ready text. Two main Document Loaders modules powered by Unstructured, the UnstructuredFileLoader and the DirectoryLoader, facilitate the conversion of different file types by intelligently identifying data formats and extracting text. This integration simplifies the process of loading diverse document types into LangChain, thereby expanding its application potential and ease of use.

Millions are using ChatGPT, but its knowledge is limited to pre-2021 data and lacks awareness of recent or private data. This blog post provides a tutorial for setting up a customized version of ChatGPT using a specific data corpus, with an accompanying GitHub repository for reference. The process involves two main components: data ingestion and creating a chatbot interface. Data ingestion includes loading data from various sources, chunking it into manageable pieces, embedding those chunks, and storing them in a vectorstore for efficient querying. The chatbot setup involves combining chat history with new questions to form standalone queries, using these to fetch relevant documents, and generating responses using a language model. The tutorial discusses customization options, such as altering prompts and selecting different language models, and offers guidance on deployment, including using a simple terminal interface or deploying via Gradio and Hugging Face spaces.

GPTwitter is a novel social media platform developed by Akash Samant and colleagues that uses Large Language Models (LLMs) to generate personalized tweets based on users' like histories, maintaining the familiar feel of their existing social media bubbles. By sampling directly from the latent space, the platform creates tweets that incorporate various elements such as length, style, and topics, and adapts future content to user preferences through a custom example selector. Despite challenges with LLM consistency, the developers found value in leveraging the models' creative capabilities, embracing hallucinations to produce diverse and novel content. The platform aims to explore the intersection of AI technology and user-level personalization, fostering an innovative approach to social media interaction.