AI Assistant

AI Assistant

The AI Assistant is the panel on the left of every project page. It is not a general chatbot dropped into the corner — it can see the project you are in, and it can read the things you have saved.

You do not need to know anything about AI to use it well. This page covers what it can do, and a few habits that make it cheaper and more accurate.

It already knows where you are

You never have to explain your situation first. The assistant is told which page you have open, so "what does this number mean?" works while you are looking at the number.

It is also given a short list of everything saved in the project — papers, structures, compounds, tool outputs, notes — and the recent runs and whether they finished. So you can ask things like:

• "What have I actually collected so far?"
• "Did my molecular dynamics run finish?"
• "Which of these files still has no downloaded file?"

When it needs the full contents of something in that list, it opens it itself.

Ask AI: pointing it at one thing

Most pages have an ✦ Ask AI button next to individual items — an evidence card, a run, a search result, a report. Clicking it tells the assistant this specific thing is what I mean and opens the chat with a suggested question.

This is almost always better than describing the item in words. "Explain this" with the card attached beats a paragraph of your own description, and it costs less.

What you can ask it to do

Explain results. "In plain language, what does this contact map actually show?" It reads RxCubed's own documentation for the tool before answering, rather than guessing.

Explain the platform. "How do I collect a structure?" or "What is the Run Results tab for?" Questions about RxCubed itself are answered from the built-in documentation, which is the authoritative source — not from the open internet.

Help you choose. On a Literature Search page it can see the results on your screen — each one's title, its search rank, its AI score, and its short summary — so "which two of these are most worth saving?" is a question it can answer properly. If it needs more than the summary for one of them, it can fetch that paper's full abstract itself.

Search the internet. When a question genuinely needs current or outside information that neither your project nor the documentation covers, it can search the web. It is deliberately reluctant to: most questions do not need it, and a search costs you money. If you want it to, just say so — "search online for recent reviews on this target".

Find a paper and save it. When it gives you a paper from the internet, it tries to include the paper's DOI — the permanent identifier. A small Save link appears next to any DOI in its answer. Click it and that paper is saved into your project as evidence, with its file fetched automatically, exactly as if you had found it through Literature Search. A View link then takes you straight to it.

If the same paper is already in your project, RxCubed recognises it and points at the copy you already have rather than saving a second one.

Write things up. It can draft a summary of what you have found, which you can then save as a note or fold into a report. You always click the button — it never writes anything into your project by itself.

What it will not do

The assistant can read your project. It cannot change it. It cannot save, delete, rename, or run anything — every one of those actions needs you to click a button.

This is a deliberate design decision, and it is worth knowing about for two reasons. It means nothing can be quietly altered by a misunderstanding. And it means that if a paper you saved contains text trying to instruct an AI, there is no path for that text to do anything at all.

It will also tell you when it does not know. If it cannot answer confidently from your project, the documentation, or a search, it is instructed to say so rather than invent a plausible answer. An assistant that says "I can't tell you that reliably" is working correctly.

Scoring: Literature Search and UniProt

Two searches can be scored by AI — Literature Search and UniProt. Both work the same way.

You give the search two kinds of input:

• Keywords — the words that should appear, the ordinary search.
• Qualities — what you actually care about, in a sentence, which no keyword could express. For example: "describes direct binding of an inhibitor to the catalytic zinc".

Each result then gets a score between 0 and 1. Keywords and Qualities count equally — half the score each. That is the point of Qualities: one strong match on something you actually care about can lift a paper above others that merely used the right words. A paper sitting third in the raw search can finish first once it is scored.

You can also set an importance on each keyword and quality, if some matter more than others.

Two things worth knowing:

• Open Details ▾ on a saved result to see the breakdown — one row per keyword and quality, each with its own score. That tells you why something ranked where it did.
• A paper with no abstract is not scored at all, and keeps its original search position. There is nothing for the AI to read, so RxCubed does not guess.

You can turn scoring off entirely with the switch on the search page. Results then keep the database's own ordering, and nothing is sent to an AI model. Your choice is remembered per search.

Two models: Luna and Sol

RxCubed offers two models. Luna is the everyday one — quick and inexpensive. Sol is the deep thinker — slower and considerably more expensive per question.

You do not have to work out which to use, because each one will tell you.

• On Luna, if a question turns out to need genuinely hard multi-step reasoning, or is about weighing up conflicting evidence, or is a consequential experimental design decision, it will answer as best it can and suggest switching to Sol. It is told not to suggest this for ordinary explanations, summaries, searches, or straightforward comparisons — so a suggestion means something.
• On Sol, once you are back to routine work — navigation, short explanations, summarising, simple comparisons — it will suggest going back to Luna. This suggestion has a deliberately low bar, because staying on Sol out of habit is the single easiest way to spend more than you need to.

Either way it is only ever a suggestion, shown alongside a real answer. Nothing switches by itself, and you can dismiss it. If you dismiss or ignore one, you will not be pestered again for a few turns.

The practical advice: stay on Luna, and switch to Sol only when Luna suggests it or you know the question is hard. Switch back afterwards.

How RxCubed keeps the cost down

Every question sends the assistant some background so it can answer usefully, and that background costs money. A few things happen automatically to keep it small:

• You are sent summaries, not whole files. The assistant gets a compact list of what is in your project and opens only the specific items it needs.
• Search results send the short AI summary, never the full abstract. If a question needs more for one paper, it fetches just that one.
• Long conversations are folded into a running summary rather than resent in full every time.
• Large documents are condensed before being used, instead of being pushed through whole.
• If something had to be left out to fit, you are told. It is never silently dropped.
• Every question is counted. Settings → Usage shows what you have used, broken down by project, so there are no surprises.

Things you can do yourself:

• Use ✦ Ask AI on the item rather than describing it. Shorter and more accurate.
• Start a new conversation for a new topic. A long thread carries its own history with it.
• Stay on Luna unless the question is genuinely hard.
• Only ask for an internet search when you need one.

If chat is unavailable

The assistant needs a model to be configured. If none is available you will be told what to do — usually adding an API key under Settings. Everything else in RxCubed works perfectly well without it; the assistant is an aid, never a requirement.

What to do next

The tutorial uses the assistant at several points on real data, which is the fastest way to get a feel for it. API keys covers setting a model up.