The research assistant

What it does

When you ask a question, the assistant searches the collective's published articles and replies with the most relevant real passage, then links you to the full piece(s). It is extractive: it quotes text that actually exists in the articles — it does not generate new prose, and it does not make things up. If it can't find a good match, it says so.

If you ask for a plain-English answer — “in simple terms,” “explain simply” — it leans toward the plainest wording it can, and for some topics it can draw on short plain-language summaries and common-question answers written for this site. Those are still just text shipped with the site: nothing is generated in your browser, and the answer is assembled where you are.

Not everything typed at a chat box is a research question. A greeting, a question about the assistant itself (“what can you tell me?”), or a request it simply can't serve — write me a poem, what's the weather, translate this — isn't searched for: it gets a direct answer about what this assistant is for, plus a few example questions, instead of three unrelated articles. Anything that names a real subject still goes to the articles.

To understand paraphrases, it uses a word-relationship map built from the articles themselves (which words tend to appear near each other), so “how does it forget?” can still reach a passage about “decay” or “stale facts.”

Why it exists

The library is large. The assistant is a faster way to find the passage and the article that answer a specific question, without reading everything — a shortcut into the material, not a replacement for it.

Where it runs

The answering happens entirely in your browser. The articles' text and the word-relationship map are ordinary files this site already serves; your questions are matched locally, and no third party is involved at any point. What is recorded, on this site's own server, is what the site is being used for — the questions asked and whether the assistant managed to answer them.

What is recorded, and what is not

Nothing sensitive and nothing that identifies you. No name, no email address, no account, no IP address, no cookie, no advertising or tracking identifier, and nothing shared with anyone else. There is no way to work out who asked a question, and no attempt is made to.

What is kept is real usage and real prompts, because that is the only honest way to make this better: the questions people actually type show where the library is thin, and the ones the assistant answers badly show where it needs fixing. Concretely, a record can contain:

• the text of the question you asked (trimmed to 300 characters);
• the answer the assistant gave you — its own reply text, quoted from these articles, so a 👎 shows what actually needs fixing;
• which article(s) the assistant showed you, if any (their catalog codes);
• your rating — a single 👍 or 👎 — or, for a suggestion, none;
• a timestamp and the page path you were on;
• where you arrived from, as a category — LinkedIn, a search engine, an email link, another site's address, or unknown. Never your own address, and never anything from the page you came from beyond that. Your browser is what reports this (document.referrer), and this site asks it for the least it can give: the Referrer-Policy served with every page means a link from another site hands over that site's address and nothing else — no path, no search terms, no query string. The category is worked out here and the address itself is discarded rather than stored;
• if you moved here from another page on this site, which page that was — so we can see which articles lead readers on to others. Only a page on this site is ever recorded here; a link from anywhere else leaves it empty;
• a random tag for the browser tab you are in, which lets several questions asked in one sitting be read as one conversation. It is not a cookie, it is discarded when you close the tab, and it cannot be connected to you, to another visit, or to anything outside this site;
• for a search, how many results came back; for a question, whether the assistant answered it well, answered it weakly, or found nothing at all.

It is stored in a log that is not publicly readable, and it is used to improve which passages and articles the assistant surfaces and to tune the word-relationship map. A 👍 does a little more: it records that this wording was answered well by that passage, so the same question asked later finds the same answer, and words readers use that the articles don't (“how does it forget?” → “decay,” “stale”) get connected to the ones they do. Only a fingerprint of the passage is kept and the answer is always re-selected from the live articles — nothing another reader saw is ever replayed to you.

Ratings are counted as opinions, not clicks: rating the same answer again from the same browser changes nothing and sends nothing, repeats of the same rating are folded together, and a question rated both ways is treated as unsettled. Leaning on either button can't tilt what the assistant tells anybody else.

In addition, a 👎 or a “Suggest this topic” click may be included in a periodic anonymous digest emailed to the site's owner — the question, the page, and the answer that missed — so the collective knows what to write or fix next. The digest carries only the fields above; it still contains no identifying information about you.

What it does not contain or do

• No IP address, no cookies, no account, no name or email — the feedback record carries none of these.
• No cross-site or cross-page tracking, and no third-party analytics or ad networks.
• No third-party or cloud AI. Your questions are not sent to any external model or service; there is no API call to OpenAI, Anthropic, Google, or anyone else.
• Nothing is sent while you simply type or read — only the anonymous record above, and only when you click a rating or “Suggest this topic.”
• Nothing you type can run. The assistant treats every question as text and nothing else — it has no way to execute what you send, and the site refuses to run any script it didn't ship. Typed instructions are answered like any other question, not obeyed.
• It works the same on phones, tablets, and desktop — the answering and privacy behavior are identical on every device.

The newsletter is a separate thing

Every page carries a “follow and subscribe” band near the bottom. That is not part of the assistant, and nothing above changes because of it. Following happens on LinkedIn and needs no address from you at all. Subscribing does, and only if you type one in and press the button: the address is held pending until you open a confirmation link, so nobody can put you on the list, and it is used for the newsletter and nothing else. Every issue ends with a one-click unsubscribe that fills your address in for you and deletes the record — from the mailing list only; it does not touch LinkedIn. The assistant never sees any of it.

In short

Answers come from this site's articles and are computed on your own machine. The only thing that ever leaves your browser is an anonymous “this answer helped / didn't” signal, and only if you choose to send it. Questions? info [at] maxresearchcollective.com.