undertheseanlp/underthesea

Underthesea - AI Assistant

What it solves

Underthesea is a comprehensive toolkit that combines an agentic AI framework with specialized Natural Language Processing (NLP) tools for the Vietnamese language. It allows developers to build AI agents that can interact with multiple LLM providers, use external tools, and process Vietnamese text with high precision.

How it works

The toolkit is split into two primary components:

  1. Agentic AI Framework: A lightweight system for creating AI agents that communicates directly with LLM APIs (OpenAI, Azure OpenAI, Anthropic Claude, and Google Gemini) using only Python's standard library. It includes a tool-calling loop, streaming capabilities, and a session management system for long-running tasks.
  2. Vietnamese NLP Suite: A collection of specialized pipelines for Vietnamese text, including word segmentation, named entity recognition (NER), sentiment analysis, and dependency parsing.

It also provides an A2A (Agent-to-Agent) server implementation to expose agents via JSON-RPC over HTTP/SSE, complete with a bundled chat UI.

Who it’s for

Developers building AI applications that require both general-purpose agentic capabilities and deep, language-specific processing for Vietnamese.

Highlights

  • Multi-Provider Support: Native integration with OpenAI, Azure, Anthropic, and Gemini.
  • Zero External Dependencies for Agents: Uses urllib and json for API communication, reducing package bloat.
  • Built-in Tools: Includes 12 default tools such as a calculator, web search, Wikipedia, and Python execution.
  • Observability: Built-in local tracing (saved to JSON) and support for Langfuse.
  • Vietnamese NLP Pipeline: Comprehensive tools for tokenization, POS tagging, NER, and sentiment analysis specifically for Vietnamese.
  • A2A Protocol: Easy deployment of agents as ASGI apps with a discoverable AgentCard.