AdalFlow

AdalFlow is an open‑source Python library that lets you assemble and automatically fine‑tune language‑model pipelines. It targets developers who need to combine chat, retrieval‑augmented generation, agents, or other LLM‑based components without writing extensive glue code.

The core of AdalFlow mirrors a PyTorch‑style API: you create objects such as Generator, Embedder, Retriever, or Agent and connect them into a workflow. Each component wraps a model service, OpenAI, Anthropic, Groq, Ollama, and several others, through optional client packages, so you only install the providers you actually use. After installing the library with `pip install -U adalflow`, you set the relevant API key in a `.env` file or pass it directly, then instantiate a generator like `adal.Generator(model_client=adal.OpenAIClient(), model_kwargs={"model":"gpt-3.5-turbo"})`. You invoke the generator by calling it with a dictionary of prompt arguments, for example `response = openai_llm(prompt_kwargs={"input_str":"What is an LLM?"})`; the call returns the raw model output, typically a string or a structured JSON if a parser is attached.

Beyond basic generation, AdalFlow includes utilities for prompt parsing, structured output, text splitting, and memory management. The Embedder component converts raw text into vector embeddings, which can be stored in a vector store and later queried by a Retriever such as FAISS, Qdrant, PostgreSQL, or BM25. A typical RAG workflow therefore takes a user query, splits it into chunks, embeds the chunks, retrieves the most relevant passages, and feeds them to a Generator along with the original prompt. Agents follow the ReAct pattern; you create an Agent object, define tool functions, and run it with a Runner that handles tool calls and response aggregation. Tracing and logging modules record each step of the pipeline, producing a trace graph that can be inspected for debugging or performance analysis.

Documentation is organized into tutorials, core concepts, and API references. The “Getting Started” guide walks you through installing the package, configuring environment variables, and issuing a first query. Additional sections cover advanced topics such as streaming responses, human‑in‑the‑loop workflows, and integration with external logging platforms like MLflow. The project is community‑driven, with contribution guidelines and a contributor guide available on the site, encouraging users to extend the library with new providers or components.

Key features

• PyTorch‑like component architecture for generators, embedders, retrievers, agents, and more