Quivr is a Python-based framework for building retrieval-augmented generation (RAG) systems that integrate generative AI capabilities into existing applications. The project positions itself as an opinionated RAG solution designed to abstract away the complexity of RAG implementation, allowing developers to focus on their product rather than infrastructure. The framework is available as quivr-core and can be installed and integrated into projects with minimal setup, requiring only Python 3.10 or newer.
The core functionality centers on flexible LLM and vector store support. Quivr works with any large language model including GPT-4, Groq, and Llama, and supports multiple vector stores such as PGVector and Faiss. The framework handles diverse file formats including PDF, TXT, and Markdown, with extensibility for custom parsers. Users can configure RAG workflows through YAML files, with a basic RAG setup requiring only five lines of code. The system supports customization through features like internet search integration and tool addition, with documentation available at core.quivr.com.
The project integrates with Megaparse, a companion tool for file ingestion, creating a complete pipeline from document processing to question-answering. The framework supports multiple API providers including Anthropic, OpenAI, and Mistral, as well as local models through Ollama. Configuration is handled through environment variables for API keys and YAML files for workflow definition.
The project is backed by Y Combinator and Theodo, providing institutional support for development. The framework emphasizes privacy and security as core topics, reflecting concerns relevant to applications handling sensitive data through AI systems.
