FieldQuest: one local AI quest to get you outside

This submission, FieldQuest, emerged from the Hacktoberfest Open-Source AI Challenge's Week 1: Touch Grass theme. FieldQuest generates brief outdoor quests based on two user inputs: available time and current mood. Upon launching, users simply read the prompt, keep their phone, and engage in noticing something. The app does not request location data.

Quests are concise, adaptable to known public spaces, and aim to keep digital devices out of the experience. The demo, code, and GitHub repository are available for exploration.

FieldQuest leverages Hugging Face Transformers.js to run the open-weight SmolLM2-360M-Instruct model directly within the browser. Utilizing a q4f16 ONNX export, the model's weight file is approximately 272 MB, alongside additional assets required for tokenization and browser runtime. Once the necessary assets are cached, the app operates offline, allowing for seamless performance.

The model generates a short nature-focused detail, which FieldQuest then transforms into a tailored quest based on the chosen time and mood. The generated phrase undergoes formatting checks and undergoes filtering against a restricted term list. While these measures provide a limited safeguard, users are encouraged to select familiar public areas, adhere to path usage, and exercise personal judgment regarding environmental conditions.

No application server is employed, with user selections pertaining to time and mood residing solely in page memory. They are only utilized for local inference, eliminating the need for GPS tracking, account creation, or analytics. The open-source nature of the model weights and inference code empowers users to examine, modify, self-host, or replace the system. Ultimately, FieldQuest provides a means to engage with nature, all while circumventing the need to transmit personal mood or location data to external services.