Backup chapter A · Materials & the Environment

Ask the Virtual Prof

A chatbot that has read Chapters 1–6. It finds the passages that answer your question, shows you where to read them, and can explain them in its own words with a small open language model that runs on your device.

1Ask a question

The Virtual Prof answers from the book and nothing else. With AI explanations off, it replies instantly with the best passages, highlighted and linked. Turn them on and a language model reads those passages and writes a short answer that cites them; the model is downloaded once from the Materials & the Environment organisation on Hugging Face and then runs in your browser, with no server and no account.

Figure A.1The Virtual Prof. Numbers in coloured boxes are citations: click them to open the section of the book they come from. The same assistant is available on every page through the Ask the prof button, where it looks first in the chapter you are reading.

2How it works

A language model small enough to run in a browser cannot remember a whole book, and when it tries it makes numbers up. The Virtual Prof therefore separates finding from explaining, a design known as retrieval-augmented generation.

Index the book

Chapters 1–6 are cut into passages of about 170 words at section boundaries, plus fact cards generated from the book's data files: materials, fibres, composites, eco-audit products.

Retrieve

Your question is matched against every passage with BM25, a classic ranking function, with synonyms (stiff → modulus, CO₂ → carbon) and a boost for the chapter you are reading.

Explain

The best four passages are numbered and given to the model with one instruction: answer only from these notes, cite them, and say so if they do not contain the answer.

Point back

Every answer ends with links to the sections it used, so the book, not the model, has the last word.

Figure A.2From question to answer. Steps 1 and 2 run instantly in any browser. Step 3 needs the model download and, ideally, WebGPU.

Retrieval is lexical: it matches words, not meanings. It is fast, transparent and needs no download, and the synonym list covers the vocabulary of the book. You can watch it work below: type a question and see the passages it would hand to the model, with their scores.

Figure A.3Inside the knowledge base. Left: the top passages for a query and their BM25 scores. Right: passages per chapter, split into text from the chapters and fact cards built from their data.

3The models

Two open models from the Qwen3 family are offered, used in their non-thinking mode, converted to ONNX and quantised to 4 bits for Transformers.js. They are published under the MaterialsandtheEnvironment organisation on Hugging Face, with the book's index alongside them as a dataset; if a model is not found there, the page falls back to the public ONNX build it was copied from.

Option	Model	Download	Best for
Fast	virtual-prof-qwen3-0.6b	≈ 0.6 GB	most laptops with WebGPU; short answers in 1–3 s, usually faithful, sometimes too confident
Better	virtual-prof-qwen3-1.7b	≈ 1.4 GB	desktops with a recent GPU and plenty of memory; clearer explanations. Experimental: it may fail to load on smaller machines, in which case use Fast

With WebGPU (recent Chrome or Edge), answers start within a second or two; on the CPU fallback, expect several seconds per sentence. The download is cached by the browser, so it happens once.

4Limits

• It only knows the book. Questions outside Chapters 1–6 get an honest “not in the notes”, or a poor match. That is by design.
• Small models still slip. A 0.6B model can misread a passage, cite the wrong note or refuse a question its notes do answer. Every claim carries a citation: check it.
• Retrieval is literal. If a question uses words the book does not, rephrase it with the book's terms (modulus, embodied energy, eco-index, reinforcement form).
• Figures are not read. The prof sees captions and text, not the interactive charts. For exact rankings, use the figure itself; the prof will tell you which one.

5From a small model to a virtual prof

An earlier version of this chapter fine-tuned a 360-million-parameter model (matenv-smollm2-360m) on 2 400 questions generated from nine calculators of Chapters 1–4. It learned to call the right calculator for about 80% of unseen questions, up from 0% for the base model, and its direct answers rose from 16% to 72% correct. It showed that a small model is a good router and a poor memory. The Virtual Prof keeps that lesson and scales it to the whole book: the memory is the book itself, searched exactly; the model only reads and explains. The training material remains in the repository's training folder.