August 2023 Summaries

Epsilla's integration with LangChain represents a significant advancement in the domain of question-answering systems by utilizing Retrieval Augmented Generation (RAG) to address the limitations of Large Language Models (LLMs) like ChatGPT. This integration enables dynamic retrieval of up-to-date information from external sources, overcoming the models' inability to incorporate knowledge beyond their last training cut-off and their lack of access to proprietary data. LangChain provides a streamlined interface for building generative AI applications, allowing developers to bypass redundant coding and focus on enhancing application value. By leveraging Epsilla's vector databases, developers can efficiently implement a question-answering pipeline that retrieves relevant documents based on semantic similarity, thereby enhancing the accuracy and relevance of LLM-generated responses. This collaboration promises to deliver richer, more context-aware answers and positions these tools at the forefront of AI-driven transformations.

A new partnership between Epsilla and CambioML aims to enhance the capabilities of machine learning scientists in fine-tuning large language models (LLMs) by integrating Epsilla's vector store with CambioML's open-source library, pykoi. This collaboration facilitates the collection of labeling demonstration data, training with reinforcement learning from human feedback (RLHF), and model comparison. A demo showcasing this integration involves creating a chatbot that can conduct personalized Q&A sessions using data from uploaded documents like books. The pykoi library offers a declarative Python interface, allowing developers to create applications with components such as chatbots and feedback dashboards. Users can try the demo themselves by installing the necessary software and accessing a user interface to interact with the chatbot, upload files, and provide feedback on response quality.

Epsilla is a high-performance open-source vector database that employs a novel parallel graph traversal technology to achieve lower query latency compared to traditional HNSW index-based systems. In a benchmark study using the gist-960-euclidean dataset, Epsilla was tested against top vector databases such as Qdrant, Weaviate, and Milvus, utilizing different setups and precision targets. The results demonstrated that Epsilla outperformed these competitors by up to 10 times in query latency and up to 5 times in query throughput within a 95% to 99% precision range. Despite striving for objectivity, the study acknowledges potential biases due to differences in parameter optimizations across databases. Exclusions included Pinecone, due to its SaaS-only offering, and non-native vector databases, as past benchmarks have shown native databases to be significantly superior. Epsilla intends to expand its testing with more datasets and optimizations, planning to open-source its benchmark code for community feedback and improvements.