AIBackends: A Framework-Agnostic AI Task Library to Free Workflows from Framework Lock-In

AIBackends is a framework-agnostic Python AI task library designed to address the pain point of framework lock-in in AI development. It offers pre-built tasks like invoice extraction, PII redaction, and document classification, supporting seamless switching across multiple frameworks such as LangGraph, pydantic-ai, and OpenAI Agents SDK—allowing core business logic to be independent of framework choices. Additionally, it supports local-first and cloud-based model execution, balancing data privacy and flexibility.

In the current AI development landscape, there are numerous frameworks, each with unique abstraction methods, configuration approaches, and invocation patterns. This leads to deep coupling between business logic and frameworks, requiring extensive code rewrites when switching frameworks. AIBackends adopts the core design philosophy of "framework agnosticism", abstracting AI tasks into independent, reusable units exposed via a unified interface, which can adapt to multiple frameworks and significantly reduce migration costs.

AIBackends is built around four core concepts:

1. Task: A functional unit directly invoked by users (e.g., extract_invoice), defining clear input/output contracts and returning structured Pydantic models;
2. Runtime: A universal LLM executor that provides complete()/embed() interfaces, supporting local (llamacpp, transformers) and cloud (anthropic, groq) runtimes;
3. Backend: A pluggable implementation of specific capabilities (e.g., PII detection supports gliner and openai-privacy backends);
4. Model: Model configuration management to simplify the complexity of model switching.

AIBackends provides a variety of ready-to-use tasks:

• Document Understanding: Summarization, classification, custom Schema extraction (including optimized invoice extraction tasks);
• Privacy & Security: PII redaction (identifying sensitive information like emails and phone numbers);
• Multimedia Analysis: Sales call analysis, video ad analysis;
• Embedding Tasks: Unified text vectorization interface. Tasks can be combined into workflows, supporting retries, step orchestration, and batch processing.

AIBackends supports offline execution of local models (via llama-cpp-python, etc.), making it suitable for privacy-sensitive scenarios. It allows downloading model caches via the aibackends pull command, optimized for NVIDIA GPU/Apple Silicon. Additionally, it enables seamless switching to cloud runtimes (anthropic, groq, etc.), allowing dynamic selection of execution environments without modifying business code.

AIBackends can integrate with various Agent frameworks:

• LangGraph: Used as a node function;
• pydantic-ai: Aligns with type safety requirements and used as a tool;
• OpenAI Agents SDK: Wrapped as a Function Calling tool;
• CrewAI/Agno: Implemented as Agent capabilities;
• Custom Applications: Directly invoked in web frameworks (Flask/FastAPI) or CLI scripts to achieve multi-framework reuse of the same task logic.

Applicable Scenarios:

1. Enterprise environments with coexisting multiple frameworks;
2. Projects requiring framework flexibility;
3. Local-first privacy-sensitive applications (healthcare, finance, etc.);
4. Batch processing ETL scenarios (invoice processing, document classification, etc.). Limitations: Focused on predefined structured tasks; complex reasoning chains need to be used in conjunction with Agent frameworks.

AIBackends provides a robust and flexible AI development solution by encapsulating stable tasks and abstracting variable choices (frameworks, models, runtimes). The project is open-source under the Apache 2.0 license and is in active development, making it suitable for teams looking to break free from framework lock-in and support local-first approaches.