AdalFlow

AdalFlow is an open-source, PyTorch-inspired library designed for building and automatically optimizing language model (LM) workflows, including chatbots, RAG systems, and agents. It provides a unified framework with components like Generator, Embedder, Retriever, Tool Use, Agent Runner, Streaming, Tracing, and Human in the Loop. The library emphasizes auto-optimization through techniques like few-shot optimization and text-grad, enabling users to improve model performance without manual prompt engineering. AdalFlow integrates with major AI providers such as OpenAI, Anthropic, Ollama, and others, and supports a wide range of retrieval backends (BM25, FAISS, LanceDB, Qdrant, etc.). It is community-driven and suitable for developers seeking a flexible, programmable approach to LLM application development, from experimentation to production. AdalFlow's PyTorch-like API gives developers fine-grained control over prompts, models, and output parsers, using Jinja2 templates and DataClasses for structured data. The library supports training and evaluation modes via a Component class, and includes built-in evaluation benchmarks (BigBench, TREC, HotpotQA). Recent updates include support for AdaL CLI, a self-evolving coding agent that uses auto-optimization for code generation. Compared to no-code platforms like LangFlow or Voiceflow, AdalFlow is strictly for programmers who want to code their LLM pipelines. It offers more flexibility than higher-level libraries like LangChain, especially for optimization tasks, but requires more setup and understanding of the underlying models. It is best suited for teams that need to iterate on prompt engineering, retrieval, and agent behavior programmatically.