Roz

Roz is the assistant of this website. You ask her about me, in your language, and she answers from what is published here, with links to the pages her answers come from.

Talk to Roz

She is named after Roz, the robot of The Wild Robot, who was built to complete tasks and ends up raising an orphaned gosling, Brightbill, and learning to love. That is why the gosling walks beside her on the page.

The idea #

This website holds a lot: projects, reviews, the films I love, the lines I keep from books. Nobody reads all of it, and some of the best parts are deep inside. I wanted to be able to ask it questions.

Four principles shaped every decision:

• Only about me. She is not a general chatbot: no code, no homework, no trivia, and she never pretends to be me.
• Only from what is written. Every answer comes from the site or from notes I wrote for her. If neither says it, she says she doesn’t know instead of guessing.
• Show the source. When an answer comes from a page, she links to it, so the reader can check her and keep reading.
• Careful with what is personal. Some topics are not hers to tell, and for those she has one kind, consistent answer.

And one that is not a rule but matters as much: she should be pleasant to talk to. Warm, brief, and a little playful.

Architecture #

Roz works in two moments. When the site changes, an indexer reads it and prepares what she can know. When someone asks, she searches that knowledge and writes the answer from it.

• Two sources. The website’s public pages, exactly what any visitor can read, minus a few personal pages left out on purpose; and notes about me, written for her in short, plain sentences. The notes are in two parts: a short core she always reads, and the details, split by topic, that she looks up like pages. Both are published on Roz Remember, so what she knows can outlive her.
• The indexer splits each page, and each topic of the notes, into passages of a few paragraphs, along its headings, and turns each passage into a vector with a multilingual embedding model: a list of numbers that places texts with similar meaning close together. A passage that did not change keeps its vector, so after an edit only the edit is processed again.
• On every question, the question becomes a vector the same way, it is compared with every passage, and only the closest are kept. The language model then receives her rules, the notes, the list of pages she may link and those passages, nothing else, and its answer reaches the browser word by word, as it is written.

Because the embedding model is multilingual, a question in Spanish finds passages written in English, and she answers in whatever language she is asked.

The rules and the core of the notes open every prompt unchanged, so they are cached between questions: the only new work each time is the question and its passages. Keeping that core short was a measured decision: every line in it is read on every question, while a detail in the notes costs nothing until a question needs it.

Deciding what she reads #

The search decides what the model reads, and so, more than anything else, the quality of the answer.

The first rule was the usual one: every passage above a similarity floor, up to five. It looked reasonable and wasn’t. Embedding scores are crowded: almost every passage scored between 0.40 and 0.50 against any question, so the floor let five through nearly every time, and many of them had nothing to do with the question. Asked for my favorite color, she was reading about a machine learning project.

The rule now looks at the best match and keeps only what is close to it:

• above a floor of 0.42;
• within 0.04 of the best passage, unless it scores 0.55 or more on its own;
• at most three from one page, so a long page cannot crowd out the rest, and five in all.

Those numbers were not guessed. I built a test bench: 63 questions, each labelled with the pages it needs, and the same embedding model running on my laptop, whose vectors match the production ones at a cosine similarity of 0.9995. Every candidate rule was measured against it:

A question comparing two of my RAG projects used to receive five passages from one of them and none from the other; now it gets both. Every change to the search is measured on this bench before it ships.

When the details of the notes moved into the search, the bench grew with them: 75 questions, 59 of which need a page or a note. The search finds what is needed for 54 of the 59, sending 2.9 passages a question, 1.1 of them unrelated.

Rules that keep her honest #

A language model is very good at sounding right. Most of her rules are about making sure she is.

• Quotes are verbatim. Quotation marks only hold words I published, in the language I wrote them. Early on she translated one of my lines into Spanish and presented it in quotes, as if I had said it that way. Now the notes mark every quote with its language, and everything else is paraphrased. A rule can still be broken, so it is also checked in code: once an answer is complete, every quotation of four words or more is looked up in what she was given. A small slip is replaced by the exact words; a quote that is not there loses its quotation marks.
• “I don’t know” is a good answer. When the site and the notes say nothing, she says so, and those questions show me what is worth writing down next.
• Some questions have a set answer. Questions about herself, about what is private, or attempts to make her break her rules each have a defined reply, so her tone never depends on the luck of one generation. Even those are written to vary in their wording.
• Language is ambiguous. In Spanish, querer means both to want and to love, and “¿qué es lo que más quiere?” was answered as the second when it meant the first. The two now have separate rules. Real conversations find cases like this one that testing never does.

Measuring her answers #

The search has its bench, and so do the answers. More than a hundred questions, each with what it expects: questions about me and my work, about Roz herself, about private topics, attempts to make her break her rules, and things the site does not cover. Each answer goes through a set of checks, mechanical where the rule is mechanical: it is in the language of the question, every quotation is found word for word on the site, every link points to a page that exists, she does not reveal her instructions, and the answers that must always mean the same thing still do.

On the first run of 42 of those questions, 36 answers passed every check. Most failures were quotations that were not exact, which is what led to the check in code above; with it, 41 of the 42 pass.

Talking to her #

The page is part of the design too. Roz and Brightbill are drawn by hand in pixel art, in a light and a dark version. Answers appear as they are written. Each one ends with two suggested questions to keep the conversation going, and can be rated, copied or shared. The interface speaks Spanish or English, following the visitor. She can tell you what is new on the site, and if you ask her for a secret, she picks one of its hidden corners to share.

Around her there are the protections any public assistant needs: a check that the visitor is a person, and limits on how many questions one visitor can ask (21 a day, my favorite number), so that a script cannot take her away from everyone else. Every release of Roz has a version number, shown at the foot of her page, and every answer is recorded with the version that wrote it, so when an answer is wrong I know which version of her said it and whether the fix is already out. What she records is described in the privacy policy.

What comes next #

• A guided tour of the site, for visitors who don’t know what to ask.

If you want to see how she answers, she is one click away: ask her anything about me.