Skip to content
This repository has been archived by the owner on Jan 3, 2024. It is now read-only.

Latest commit

 

History

History
152 lines (91 loc) · 5.08 KB

README.md

File metadata and controls

152 lines (91 loc) · 5.08 KB
Raw

NPC

This is a demo using a language model agent to play text adventure games like Zork.

Contents

  • Installation
    • Build from source
  • Play game
  • Run agent
  • Custom agent scripts

Installation

The demo requires a game file for Zork. You can download the game file from the Internet Archive.

You will also need an API key from OpenAI.

Run with Docker

To run the demo with Docker, you need Docker.

To install, clone the repo and enter the npc directory:

git clone https://github.com/deepfates/npc.git
cd npc

Move the zork1.z5 file into the npc directory.

Next, create a .env file in the npc directory and add

OPENAI_API_KEY=<your openai api key>

(without the < and > marks)

Finally, build the docker image:

docker build -t npc .

To run the demo, run:

docker run --network="host" -t npc

Then open localhost:8080 in your browser.

Build from source

To build the demo, you need a version of Python 3.8 or 3.9 and Poetry.

To install, clone the repo and enter the npc directory:

git clone https://github.com/deepfates/npc.git
cd npc

Then install the python dependencies:

poetry install

Next, create a .env file in the npc directory and add

OPENAI_API_KEY=<your openai api key>

(without the < and > marks)

Finally, copy the zork1.z5 file into the npc directory.

Usage

This demo is a fully interactive text adventure game. You can play the game yourself or run an agent to suggest commands or play the game for you.

Play game

To play the game, enter the npc directory and run the server:

poetry run python server.py

Then open localhost:8080 in your browser.

Run agent

While playing the game, you can press the ⇥ button to activate the agent. The agent's internal thoughts will be displayed above the command box, and a suggestion will be displayed in the command box. You can press ⇥ again to send the suggestion and activate the agent automatically.

Custom agent scripts

The agent runs an internal loop on a prompt template. You can access this prompt template by pressing the 📜 button in the game UI. You can edit the prompt template and press the ↵ button to activate the agent with your custom prompt template. The prompt template must contain the same {placeholders} as the default prompt template.

Developing

Backend and LLM agent

To develop on the backend, you need a version of Python 3.8 or 3.9 and Poetry. Install the python environment as in the installation instructions, but run the server with the --debug flag like so:

poetry run python server.py --debug

While running server.py, your changes to the backend will be reflected in the app at localhost:8080. The server will restart each time you save edits, so if you're running the frontend you'll need to refresh the page to get a new session ID.

LLM agent

The agent currently runs a modified ReAct loop. See the ReAct paper at https://arxiv.org/pdf/2210.03629.pdf for more details.

The agent is in the npc/agent.py file. It wraps the ReAct loop with a short buffer memory and a custom prompt template. The original functionality was to run a loop for a dozen iterations or so and let it interact with the pworld through a Play tool. This is currently turned off for the purposes of the demo, because the Play actions were invisible to the frontend. Now the agent takes up to 3 actions including Look and Inventory, then recommends a command. Prompt tuning is yet needed for this functionality to work well.

The agent is built with LangChain. See their excellent documentation for more on the underlying functionality.

Frontend

To develop on the frontend, you need Node.js and NPM.

To install the frontend dependencies, enter the client directory and run:

npm install

To run the frontend in development mode, run:

npm run dev

To build the frontend for production, run:

npm run build

Once the frontend is built, or while in develop mode, the app can be served by running the server as in the installation instructions. Your changes to the frontend will be reflected in the app at localhost:8080.

The frontend is built using Svelte. The main API calls and the input form are in the App.svelte file. The components directory contains display components.

Svelte compiles to javascript, so once the frontend is built, you can run the server without the frontend dependencies. The Flask backend will serve the frontend files from the client/build directory.

The API functions in the main Svelte app are not very idiomatic. I think they should incorporte reactive assignment better. But they do the job for now.