Building a Conversational Game NPC

I experimented with creating an intelligent conversational NPC powered by a large language model. It has a distinct character identity, responds to natural-language input from the user, and plays animations that match the emotion of its reply. The NPC also has a memory system: it can remember personal information about the user and answer accurately when asked about those specific details.

For this project, I built the backend with Python's FastAPI library and used LangChain to call the language model. Memories are stored in a local vector database. Unity serves as the frontend, handling user input and 3D rendering.

Below are some key code excerpts and demonstrations running in Unity:

Calling DeepSeek-V3.2 through its API.

Python code reading an API key from an environment variable and configuring a DeepSeek model call — Calling the language model01

The system prompt gives the language model its character identity.

A Chinese system prompt describing the NPC character identity and personality — Defining the character02

Building the local vector database with ChromaDB.

Python code loading local text and splitting it into overlapping chunks — Preparing local memory documents03

The system prompt strictly requires the model to include an "emotion tag" with every response. The C# code in Unity reads this tag and plays the corresponding animation.

The Unity Animator.

The Unity frontend code.

Unity C# code handling player input and sending a dialogue request — The Unity frontend05

Demonstration.

A character responding through a dialogue interface in Unity — A conversation in Unity06

Testing the memory system.

The Unity dialogue interface showing a memory recall test — Testing memory recall07