Case Studies
Connect external data to LLMs, no matter the source.
"Carbon is truly a time saver to integrate Retrieval Augmented Generation to any AI app. I used it for the Training Data feature in Typing Mind Custom and it works perfectly. The Carbon team is also super nice and eager to listen to feedback and suggestions for new features. Highly recommended."
Tony Dinh
CEO, TypingMind
Challenge
TypingMind has been used by hundreds of thousands of users as it provides a 10x better user interface for ChatGPT. Back in the summer of 2023, TypingMind was looking to launch TypingMind Custom, a private ChatGPT for teams that allow customers to build private and secure internal AI chat apps.
TypingMind Custom previously had limitations in terms of training data as users could only upload local documents, which was constrained by the model's context length. Moreover, the uploaded documents were not saved after the chat session ended.
Solution
Carbon improved the user experience by enabling users to upload documents in multiple formats such as CSV, PDF, and DOCX from cloud sources like Google Drive, Notion, and OneDrive. Additionally, Carbon converts the uploaded documents into vector embeddings and stores them in a vector database to be referenced by models anytime across all sessions.
Here's a breakdown of several Carbon products leveraged by TypingMind Custom in their integration:
Prebuilt React Component ⚙️
TypingMind uses our pre-built React component which could be installed in a React application within minutes. This provided a user interface allowing users to authenticate and upload content from various enabled connectors, such as Google Drive, Notion, OneDrive, etc as training data. The React component also has ready-to-use screens for users to upload local files and submit URLs for web scraping.
Custom Tags 🏷️
TypingMind uses our custom tags and advanced filtering functionality. Users can add tags to documents uploaded via Carbon and assign these tags to certain AI Characters. These AI characters can access the training documents linked with those tags.
Semantic Search 🔍
TypingMind uses our built-in, lightning-fast semantic search to retrieve relevant content from uploaded documents. The search results are then incorporated as training data for models. The system automatically updates the document index as new documents are added or modified.
Conclusion
Our collaboration with TypingMind has evolved into a strong and dynamic partnership, with ongoing exploration of opportunities for expansion. We are actively working on integrating our hybrid search functionality with TypingMind Custom, enabling them to conduct more precise searches across all files and data sources for relevant content.
