Projects and Studio

Open Projects in the sidebar and choose New project: a name, and optionally a description and instructions. A project has five sections: Chats, Sources, Studio, Instructions and Memory, each at its own address. Every chat inside the project follows its instructions and can search all of its sources.

Memory shows the project’s memory files, the decisions and findings RBAi keeps for its chats (open one to edit it), and what RBAi keeps from the chats: a summary, the genes, proteins, diseases, drugs and pathways they named, and a short note of each answer. RBAi uses them in the project’s next chats; forget any of them, or all, on that page.

Add sources, in Sources and in a project chat’s source picker, opens one dialog. Drop several files on it at once, or choose a card for the other kinds:

• a PDF, Word, Excel, OpenDocument spreadsheet or PowerPoint file, up to 25 MB;
• Recording: MP3, M4A, WAV, OGG, WebM or FLAC (up to 25 MB; M4A, OGG, WebM and FLAC up to 3 MB), or a link to one, transcribed and cited by time, for example S3 4:05;
• Paper: papers by DOI, PMID, PMCID or arXiv ID, one per line;
• Web page: https links, one per line (each page is opened in a browser, so pages built with JavaScript are read too);
• Paste text: a title and the text;
• a PDF or Word file already attached in a project chat, with Add to project.

Each item has its own row: the upload’s progress, then Being read, then Ready to cite, or the reason it failed with Retry. Up to three sources are read at a time; the others wait their turn, and are added only while the page stays open. The dialog’s footer shows how many sources the project has of your plan’s limit. A paper’s full text is read when it is open (arXiv, PubMed Central or the publisher); otherwise its abstract is, and Sources says which.

Scanned PDFs need text recognition (OCR), which is not supported yet. YouTube links cannot be added: YouTube shares a video’s captions only with its owner.

Once a source is ready, RBAi writes its source guide: a short summary and its key topics, shown under it in Sources and in the source picker.

A project chat cites the passages it used as chips such as S2 p.7, and a slide as one slide. Point at a chip to read the passage; click it to open the PDF at that page with the passage marked in yellow.

Search matches meaning and exact words together, so a gene name, a number or a model name finds the passage that holds it. A phrase in double quotes must appear word for word, for example: What does the paper report for "label smoothing"?

All sources, next to Thinking in the composer, chooses which sources a question uses.

Each project has its own computer in RBAi’s cloud, and every chat of the project, on the website and in the CLI, runs its code there: Python, the shell and background jobs. Files, results and the software a chat installs stay on it between chats and days, so a new chat can use what an earlier one downloaded or installed instead of doing it again. Chats outside projects keep their own sandbox, as before.

It starts by itself the first time a chat in the project runs code, and pauses a minute after its last use. Paused, it keeps its files, its memory and its running programs, uses no code sandbox time, and resumes in about a second when a chat needs it. A background job keeps it running until the job ends, with the tab closed.

A computer unused for 90 days is deleted, with the files and software on it. 14 days before, RBAi tells you in the bell and by email, and the card shows the date; any code that runs in one of the project’s chats keeps it. The project, its chats and its sources stay, and a new computer starts the next time code runs in it.

On the project’s Chats page, the Cloud computer card shows whether it is running or paused and when it was last used, its size and disk, the background jobs on it, and how many project computers your plan keeps. Pause now pauses it, and its jobs with it. In its … menu, Restart from its disk keeps the files but stops the programs running on it, for a computer that seems stuck, and Delete the computer deletes its files and installed software; a new one is made the next time a chat runs code. Deleting the project deletes its computer.

Kept means running or paused. When your plan keeps its number already, a chat in another project can’t run code until you delete one of them or change plan. When it runs its number at once, the least recently used one without a background job pauses to make room. After a change to a smaller plan nothing is deleted: the most recently used computers keep working, and the card of each other one says how many to delete. A computer’s running time, the minute before it pauses included, counts as your code sandbox time (Plans); paused time is free.

On the project’s Chats page, Where this project’s work runs sets it for all of the project’s chats, on every browser and device you sign in on.

RBAi cloud computer

The default. Code runs on the project’s cloud computer. Your project’s cloud computer

My computer

Code and files stay in a folder on your computer, through rbai bridge. Nothing runs or is uploaded in RBAi’s cloud from the project’s chats. When the computer is offline, the chat says so, and the work does not move to the cloud.

Both

The folder for your computer’s files and programs, and the cloud computer for heavy work.

My computer and Both need a folder: start rbai bridge in it, link this browser once (The RBAi bridge), then choose the folder there or with the folder button in a chat’s message box. Choosing a folder in a chat of a project on its cloud computer makes the project Both, and Stop using this folder makes it the cloud computer again. With My computer, RBAi still looks up databases and the literature for you; only code and files stay on your computer.

Open a project and choose Studio. One chat bar asks What would you like to make from these sources? Describe the output in your own words and press Enter, for example A 20-minute journal club for specialists, dark theme or A graphical abstract for Cell Press. The bar has the Thinking control and the microphone, as in the main chat, and a sources control, such as All 4 sources, that chooses which ready sources the output uses.

Chips above the field choose the kind yourself: Presentation, Visual summary, Briefing doc and Agreements and contradictions, then your saved templates. A chip changes the examples in the field, and pressing it again turns it off. With no chip, RBAi reads the request first (Reading your request) and picks the kind and its details. A request it cannot read makes nothing and says so, with Try again.

Presentation

A talk as an editable PowerPoint file, built live in the editor. Presentations

Visual summary (Beta)

One picture of the sources’ story, planned, drawn and reviewed by AI. Visual summary

Briefing doc

A summary, the key themes, the findings, the methods and the limitations the sources state.

Agreements and contradictions

Findings that two or more sources share or dispute, each side in its own words. It needs at least two ready sources.

Anything else

A custom output: a document, a table with the columns you name, or a list, such as an evidence table or a timeline. Custom outputs and skills

A line confirms the start, and the output is at the top of Created in this project. Each output uses credits for the model calls it makes. Outputs made earlier as evidence tables, methods, rigor checklists or timelines still open as before; to make one now, ask for it in the bar.

The AI writes each value from the passages of its source and names the passage it comes from, with the words it rests on. RBAi finds those words in the source, so the value’s chip shows them and opens the source there; when the words are not found, the chip opens the passage the AI named. Check a value against its source before you use it.

A missing value says what was read: Not reported in this source when the whole source was read, Not found in the part read for a long source read in part, and Not reported in the abstract for a paper with no open full text. A source of up to 100,000 characters is read whole; a longer one is read in evenly spaced parts, and the output says how much was read.

Briefing docs and custom outputs are cited the same way: each statement, item and cell names its passages, and a sentence with no citation to a source passage is left out, with a line that says how many were. Chips such as S1 p.3 follow each value: point at one to read the source’s words, and click it to open the source at that place.

Anything that is not one of the four kinds is a custom output, written as a Document, a Table with one row per source and the columns you name, or a List. RBAi takes the form and a short title from your request. It is cited like the other kinds. When you ask for something the sources cannot support, such as a budget or a prediction, the output holds what they do support, and one line at the top says what could not be done and why.

+ in the bar attaches a skill, a written procedure for the output: Attach a skill file (.md, .txt) or Paste a skill. The skill’s text, then your request, become the output’s instructions, up to 100,000 characters.

After a custom output starts, Save as my template keeps it as a chip in the bar, in all your projects, and × on the chip deletes it. You can keep up to 50.

Presentation builds a research talk as a PowerPoint file (.pptx), live, in a full PowerPoint editor in the browser. Say in the request what the talk is. RBAi reads these from it, and chooses what you leave out:

• the kind of talk: a journal club (usually 20 minutes), a lab meeting (15), a conference talk (12) or a thesis defense (45);
• its length, from 5 to 60 minutes, at about one slide a minute;
• the audience: specialists or a mixed audience;
• the theme: light, dark or neutral.

Everything else you write, such as the slides to include or the order of the story, goes to the AI as you wrote it.

The presentation opens at once in the editor, and the AI builds it there with the editor’s own tools, from the sources’ text: the theme, the slides, text, shapes, tables, native charts of the sources’ numbers, pictures it draws, and speaker notes with each slide’s sources. You watch each slide appear, and the panel beside the editor lists every step in plain words. The AI thinks at the Thinking level in the panel’s composer, the same level as in the main chat.

It keeps going if you close the reader. A long build continues past 15 minutes in a new run on its own; the panel shows Continuing in a moment in between. Stop ends the run and keeps what is built as a version.

Then the presentation is its PowerPoint file. On a tablet or a computer (768 px wide or more), change it in the editor as in PowerPoint (text, shapes, tables, charts, slides, notes, themes) and Save a new version; on a phone the slides are read-only and the AI panel works. To ask the AI, write in the panel: the slide you are on, and what you selected, go with the request, and its changes land live. Ask for a picture, for example Add an illustration of MDM2 binding p53 on the right half of this slide, and the AI draws it onto the slide with its alt text. Present shows the slides full screen. Each AI run and each Save is a new version, and the version menu has Make version N current.

Your plan sets how many presentations you can make a month; failed ones do not count, and one more costs credits once they are used. Each picture the AI draws counts as one Visual summary drawing; past the month’s drawings it costs 175 credits, charged only once the picture is on the slide. When your credits run out, the AI stops, keeps what it built as a version and offers more credits. What each thing costs

Visual summary draws one picture of the sources’ story: a journal’s graphical abstract, a slide, a poster panel or a social post. Name the format in the request, for example a graphical abstract for Cell Press, or let the AI choose one of these:

The AI reads the sources and plans the picture: its story, its labels and the prompt for an image model, which draws it whole, words included. The AI then looks at the drawing against its plan. When something matters, such as a misspelled label or an arrow that says the opposite of the story, it has the picture drawn again with those changes, at most twice. Its row shows each step.

Edit with AI, on the picture’s top-right corner, opens a chat beside the picture (under it on a screen narrower than 1024 px), as for a plot or a structure in a chat. Describe a change and press Enter: the AI replies with what the new version changes, and when its review did not accept a drawing, a line under the reply says what it still asks for. Each change draws a new version from the one shown and keeps the rest of it. Choose any version in the versions menu and restore it, download it as PNG or PDF, or Insert into Report Writer.

Your plan includes a number of drawings a month: a new picture and an Edit with AI each count as one, its redraws included, and the viewer shows how many you have used. Past them, each picture the image model draws costs 175 credits. A run holds 525 credits when it starts (three pictures, the most one run draws), charges the pictures it drew and gives the rest back; a failed run gives back all of it. Planning and reviewing use credits for their tokens. A finished literature review has a Visual summary button too: it opens the same viewer, with the review’s newest visual summary or the formats to draw one. What each thing costs

An output opens full screen. The header has its title, kind, sources and age, the version menu and its actions; the details column beside it can be collapsed. On a phone the actions sit in a bar at the bottom, and Esc or Back closes the output.

• A presentation opens in the PowerPoint editor, with its ribbon and slide thumbnails, and the AI panel beside it (below it on a screen narrower than 1024 px; on a phone the slides are read-only). Its main action is Download .pptx, or Save while you have unsaved changes.
• A visual summary fills the window. Edit with AI sits on the picture’s top-right corner, and beside it, when you point at the picture, a switch shows the picture at its actual size.
• A briefing, timeline, custom document or list reads in one column, and a table uses the window’s width. A table’s main action is its CSV; a text’s is Edit and publish, which opens it as a new Report Writer page. Copy as Markdown and Download Markdown are there too.

Created in this project lists every output, newest first. A running row shows the step it is on, such as Reading the sources’ passages, 2 of 3 sources done, or Building the slides with the AI’s latest step, and the time since you asked. Studio keeps working when you close the tab: when you come back, the row shows its current step or the finished output.

A failed row says why, with Try again. An output that makes no progress for ten minutes is shown as stopped, with Try again.