DevShin00/furfriends-rag-v2

FurFriends RAG v2

The pet-care assistant model of FurFriends, a pet health records app made as a student thesis project (Philippines, 2026). It runs in the owner's browser with Transformers.js and writes answers only from the checked passages the app gives it: FurFriends' first-aid guides, the team's checked answers, Pets Stack Exchange answers and CDC/FDA pages. When the passages don't answer the question, it says so instead of guessing.

It is not a vet and does not diagnose. It is a study prototype, tested on 31 held-out questions graded by the student team, not by a veterinarian.

Model

• Base: Qwen/Qwen2.5-0.5B-Instruct (Apache 2.0)
• Fine-tuning: LoRA on the attention layers, 1 epoch over 813 examples, on a laptop CPU
• File: onnx/model_q4.onnx: 4-bit weights (blocks of 32) with full-precision maths, about 430 MB. Runs on WebGPU or the CPU (WASM).

Training data

Each example is an owner's question, two passages from the app's library and the answer written from the right passage (or "I couldn't find this in FurFriends' checked sources" when neither answers it).

• FurFriends' own checked answers and first-aid guides, written by the team
• Pets Stack Exchange answers from the public data dump, licensed CC BY-SA 4.0. Credit to the authors on Pets Stack Exchange. The app shows the source under every answer.
• Public-domain pages from the US CDC (Healthy Pets) and FDA (Animal Health Literacy)

Use

import { pipeline } from "@huggingface/transformers";
const generator = await pipeline("text-generation", "DevShin00/furfriends-rag-v2", { dtype: "q4", device: "webgpu" });

The prompt (system instructions and how the passages are given) is in the FurFriends repository: src/lib/aiPrompt.json and ai-server/prompting.py.