undertheseanlp/underthesea

What it solves

Underthesea provides a unified toolkit for building AI agents and performing Natural Language Processing (NLP) specifically tailored for the Vietnamese language. It simplifies the process of integrating multiple LLM providers and implementing complex agentic workflows like tool calling and multi-session task management without requiring heavy external SDKs.

How it works

The project is split into two primary components:

1. Agentic AI Toolkit: A lightweight framework that communicates with LLM APIs (OpenAI, Azure OpenAI, Anthropic, and Google Gemini) using only Python's standard library. It includes a tool-calling loop, a session manager for long-running tasks with state persistence, and a tracing system for observability. It also supports the A2A protocol to expose agents as JSON-RPC services over HTTP/SSE.
2. Vietnamese NLP Suite: A collection of specialized pipelines for processing Vietnamese text, including word segmentation, POS tagging, named entity recognition (NER), sentiment analysis, and translation.

Who it’s for

Developers building AI applications that require Vietnamese language support or those looking for a lightweight, provider-agnostic framework to deploy AI agents with built-in tool calling and tracing.

Highlights

• Zero-Dependency Agent Core: Communicates with major LLM providers using only urllib and json.
• Built-in Tooling: Comes with 12 default tools (e.g., web search, python exec, calculator) and easy custom tool integration.
• Comprehensive Vietnamese NLP: Dedicated functions for segmentation, diacritics restoration, and dependency parsing for Vietnamese.
• A2A Protocol Support: Ability to serve agents via a raw ASGI app with an optional bundled chat UI.
• Observability: Built-in local tracing to JSON files and integration with Langfuse.