WikiSpeedRunner

This project plays the Wikipedia Speedrun game automatically.
The goal: reach a target Wikipedia page from a starting page using only in-page links - no search, no URL edits, no going back, no cheating.

Two play modes:

• WikiSpeedRunner → SBERT Vector mode
• AI WikiSpeedRunner → Gemini LLM mode

---

Game Modes

Vector Mode - WikiSpeedRunner (SBERT)

Uses the SBERT model to compute word embeddings between:

every possible next-page title and the target page’s summary

It picks whichever title is semantically closest to the target.

Pros

• Very fast (no external API calls)
• Works offline after model download

Cons

• Sometimes takes more steps (longer path), even though computation is fast

Fast execution, longer routes

---

LLM Mode - AI WikiSpeedRunner (Gemini)

Uses Google Gemini to logically determine which link will lead closer to the goal using a one-line target description.

Pros

• Often picks shortest or smartest paths
• Understands concepts contextually

Cons

• Slower (each decision triggers an API call)

Short path, slower execution

---

Browser Mode (Visual Game UI)

You can enable or disable the visual browser from main.py:

USE_BROWSER = True   # show the game visually in chromium
# or
USE_BROWSER = False  # run silently (no UI)

When enabled, the bot:

• scrolls to the chosen link
• highlights it in red
• clicks it like a real player

This transforms the script into a playable AI game.

---

Installation

1️. Clone the Repo

git clone https://github.com/ankushhKapoor/WikiSpeedRunner.git
cd WikiSpeedRunner

---

2️. Install Dependencies

Using uv (recommended)

uv sync
playwright install chromium

Using pip (without uv)

pip install wikipedia-api sentence-transformers numpy python-dotenv google-genai playwright
playwright install chromium

---

3️. Environment Variables (LLM mode only)

Create .env:

API_KEY=your_gemini_api_key_here

• See /.env.example in the repository for reference
• Not required for SBERT Vector mode

---

Running the Game

Edit settings in main.py:

START_PAGE = "Dog"
TARGET_PAGE = "Linux"
PLAYER_TYPE = "vector"  # or "llm"
USE_BROWSER = True

Target Description Requirement

• If using LLM mode → set:

TARGET_DESC = "Linux is an open-source operating system..."

• If using Vector/SBERT mode → leave TARGET_DESC empty or ignore it

Then run:

python main.py

• PLAYER_TYPE = "vector" → WikiSpeedRunner
• PLAYER_TYPE = "llm" → AI WikiSpeedRunner

---

History

Each run is saved to history.json, including:

• path taken
• steps
• time taken
• player type used

Compare runs to see whether speed (SBERT) or efficiency (LLM) wins.

---

Inspiration

This project is inspired by a Green Code YouTube video I Forced AI to Speedrun Wikipedia demonstrating an AI solving Wikipedia speedruns.