AdalFlow

AdalFlow is an open source Python library for building and automatically optimizing LLM applications. It provides the building blocks for generators, embedders, retrievers, tool use, agents and structured output, and adds a training layer that tunes prompts and few shot examples against a dataset, so a question answering, classification or RAG pipeline can be improved systematically rather than by manual prompt editing. It is aimed at machine learning engineers and developers who want a lightweight, model agnostic framework with a clear component design. Model clients cover OpenAI, Anthropic, Azure AI, Bedrock, Cohere, DeepSeek, Fireworks, Google, Groq, Mistral, Ollama, SambaNova, Together, xAI and local Transformers models. Retrievers include BM25, FAISS, LanceDB, Postgres, Qdrant, rerankers and LLM based retrieval. The agent layer offers a ReAct agent and a runner with streaming, tracing and human in the loop support, plus MCP tools and a function tool abstraction. Evaluation utilities include answer match accuracy, retriever recall and LLM as judge metrics, and built in datasets such as HotpotQA, TREC and Big Bench Hard support benchmarking. Trace graphs visualize how text based gradients and demonstrations flow through a pipeline during optimization. AdalFlow is free to use and distributed as an open source package with its code on GitHub. There is no paid plan or hosted service. The documentation site doubles as the product home, with tutorials, use cases and developer notes on the library's design philosophy and class hierarchy.