It is not a Q&A bot. It is more like a personal assistant who reads what you upload, finds papers for you, drafts reports, edits Word documents, sends scheduled emails, and remembers your preferences. This article walks through what it can do and how to get the best results, using real scenarios.

Most AI assistants only "chat" — you ask, they answer. The General Assistant in aurora agent is different. It actively searches for you, reads for you, organizes for you, produces files for you, and even runs long jobs in the background. This article is not a manual. It walks through, one by one, the things it can do for you, so you know what you can confidently hand off, and what you are still better off doing yourself . What role can it play Think of the General Assistant as a personal research assistant who just joined your team: You give it materials related to your research (papers, reports, notes), and it actually reads them , not just keyword-matches; You ask it questions, and it first looks for evidence in your materials before answering. If it cannot find the answer, it tells you honestly , instead of making things up; You ask it to do something, and it comes back with the result — not a suggestion of "you could try this," but a finished deliverable; The longer you work with it, the better it knows you — your research direction, citation habits, preferred language, what you have been working on lately; For anything important or outward-facing (sending emails, deleting files, scheduling tasks), it confirms with you first , instead of acting on its own. It thinks of itself this way: "You are not a chatbot. You are a thoughtful person who happens to be very good at reading documents and finding information." Get started in three minutes Four steps gets you through one full round trip: Step 1: Create a "Research Space" In the top navigation, go to Research Space -> click New Research Space in the upper right. A Research Space is your private bookshelf . Anything you upload there is visible only to you. We recommend organizing by topic — for example, separate spaces for "AI Alignment," "Macroeconomics," and "Personal Notes." They stay isolated from each other, and the assistant will not be distracted by unrelated material when you ask questions. Step 2: Drop a few documents in Drag in PDFs, Word, Excel, HTML pages saved from the web, Markdown notes — whatever you have. The system processes them automatically: Ordinary PDFs: parsed in 30 to 90 seconds; Scanned PDFs (photos, photocopies): the assistant automatically reads the text in the images, more accurately than typical OCR; Word / PowerPoint / Excel: parsed normally; Large files (>30MB): they go into a background queue, and your browser will notify you when they are done. You do not have to read the whole book or paper yourself first. That is its job. Step 3: Start asking Switch to the chat page, choose General Assistant , and your first message can be something like: You will see what it does: it searches your materials, opens the relevant pages and reads them carefully, then gives you the answer plus a quoted block from the source (with the page number). The whole process is visible in the side panel — it is not a black box . You always know where the answer came from. Step 4: Turn the result into something you can use Then ask: When it is done, a Word download link appears right in the conversation. No copy-pasting, no opening Office on the side. Congratulations — you have just completed the basic flow. Now let's get into the things it is genuinely good at, but you might not have realized yet . What it is genuinely good at Each of the sections below is a real scenario, paired with an example of what you might say to it. You do not have to memorize any commands. The more specific you are, the better it does. Scenario 1: Read an entire Research Space cover to cover Many people assume a knowledge base is just "find a few keyword-matching sentences." Not so — the General Assistant can run the same deep analysis across your whole Research Space, or across a batch of files : It reads through every paper in parallel , and even if the network drops in the middle, it can resume from where it left off , so you do not waste half an hour only to find out nothing came back. You get a structured report at the end, plus a Word file. What is even better — once it has read them, this information stays in its head . You can follow up with "Which ones used RLHF?" or "List every passage that mentions Constitutional AI," and it does not have to read them again . Scenario 2: Track the latest research and get a daily digest This goes well beyond Google. The General Assistant has built-in access to five academic sources: PubMed (medicine and biology), arXiv (computer science, AI, math, physics), OpenAlex , Semantic Scholar , and Crossref . Once it is set up, you will get an email every day at 9 AM . You do not have to remember you set this up — it remembers for you. Want to pause it, change the topic, or change the frequency? Just say so — No need to dig through a settings page for the right toggle. Scenario 3: What happened today — a cross-source industry briefing Sometimes you do not want a literature review. You want to know what happened in the field today — product launches, technical blog posts, GitHub projects, new papers, all in one place: It will search multiple sources at the same time , then organize the results into a clean briefing with sections like "Product News / Technical Blogs / Open-Source Projects / Latest Papers." One sentence in, one finished document out. Scenario 4: Turn a stack of papers into a literature review When the request is "write a literature review based on these papers," the General Assistant switches into literature-review writing mode and follows a strict sequence: 1. Check the state of your materials; 2. Stream out questions to clarify your writing requirements (Chinese or English? Which citation style? Which angles to focus on?); 3. In parallel , extract the key points of every paper; 4. Generate an outline and pause for you to confirm — you can ask to change a section, add one, or drop one; 5. Once the outline is locked in, it streams the chapters out one by one. You can read each paragraph as it appears. Done is not the end of the story. You can then say, "In Section 3, second paragraph, change PPO to DPO," and it will edit that exact sentence and automatically save a version snapshot. Don't like the change? One sentence and it rolls back. Scenario 5: Edit your existing Word document or long draft in place (the most underrated capability) You upload one of your own .docx files, or take a long draft it just wrote in the previous turn, and you say "expand Chapter 3" or "change every 'PPO' to 'DPO'" — It does not rewrite the document from scratch . It finds the exact spots in the original and edits in place : Every edit automatically creates a version snapshot; If you don't like it, one sentence rolls it back; When multiple edits happen at once, it is a single atomic commit (either all of them succeed, or all of them roll back), so the document never ends up half-changed and broken; When it finishes, you get a Word download right there. Rolling back is free — so you can let it edit without anxiety. Scenario 6: It can compute, plot, and produce charts It does not just talk about your data — it actually runs the code for you: The font issues you have probably wrestled with on your own computer when using matplotlib for Chinese charts? Already configured in advance — the chart comes out clean, and the title does not turn into squares . Need a diagram, a flowchart, or a cover image? It will generate it for you with AI. If you have an existing image you want redrawn, composited, or annotated, just send it in. Scenario 7: Recognize text from receipts, contracts, and handwritten notes You snap a photo of a receipt, or you have a scanned contract PDF, and you want the text out — It uses dedicated image text recognition (which preserves table structure and formulas ), and if that fails, it automatically falls back to a vision model. You don't have to wonder "is this a scan? do I need a different tool?" — it figures that out itself. Scenario 8: A whole family of scheduled tasks Daily digests are just the start. Anything recurring can be put on a schedule: Every morning, organize yesterday's tracked papers into a briefing and email it to you; Every Monday, summarize what was added to your Research Space in the past week; The first of every month, give you a "research progress self-check" for the past month. Forgot what you scheduled? Just ask, "What scheduled tasks do I have right now?" and it will list them for you. Scenario 9: It actually remembers you The General Assistant does not start from zero in every conversation . It naturally remembers: Your research direction, institution, email; Your preferred response style (detailed or concise? Chinese or English? GB/T 7714 (a Chinese national citation standard) or APA?); What you have been working on recently; Important conversations you have had with it (e.g. "we discussed privacy budgets last week"); The people in your circle ("Alice is my advisor"); The lessons you have learned along the way ("last time my SQL used LIMIT and missed pagination"). At the start of every conversation, it automatically pulls up these memories — you do not have to repeat "I use GB/T 7714" every time. But it is also careful: if you previously said your research direction was NLP and later you mention CV, it will not silently overwrite the old memory . It will ask, "Is this a new direction, or are you correcting the previous one?" — this prevents bad memory from poisoning future conversations. You can also ask it to "forget that one" any time — just say so. Scenario 10: Hand off long jobs and go do something else Some tasks take a while — for instance, asking it to analyze 200 papers in your Research Space. It will arrange a background helper to run the job, and you don't have to wait . You can carry on with another conversation, close the browser, come back the next day — When you come back, it will proactively tell you "that task you started is done," and hand you the result. Halfway through, want to know "which paper is it on now?" — just ask. After it finishes, want to add an angle? "For that batch of papers we just did, can you also look at what their experimental setups have in common?" — it answers with the full original record in hand , no need to redo the work. This is the biggest difference between it and a chatbot: it can run things in parallel like a real assistant you hired , instead of one question, one answer. Some prompt templates that work well You don't need to copy these word for word — they are just patterns to riff on: Read one paper deeply: Sweep the whole library: Track new papers: Revise a long draft: Let it edit something with peace of mind: What it cannot do, and what you should not hand to it Clear limits matter just as much: Reproducing an entire paper / writing a project-scale codebase / multi-script empirical modeling — for that kind of deep engineering work, switch to the Research Paper Agent in the input box at the bottom, and upload your data and papers. The General Assistant will tell you so itself, and will not push past its limits . Production operations / long model training runs / large-database queries — beyond what it should be doing. Deleting important files / emailing strangers / publishing content publicly — it always confirms with you first . Take that step seriously; it is not just a formality. Facts that are not in your materials, it will not invent — if it cannot find evidence, it will tell you "there is no relevant material in your Research Space." That is by design, not a bug. A few tips for getting the most out of it Use your Research Space as long-term memory : throw every paper you have read in there. Six months later, you can come back and ask questions, and it can still cite them precisely. Don't pile on too much in one message : split a complex task into two or three turns. Have it do step one, show you the result, then continue. That works far better than stuffing five requirements into a single sentence. Give it "style examples" : when writing a literature review, include two paragraphs you wrote previously. It can mimic your voice. Let it edit your documents without fear : version snapshots are free. If you don't like it, one sentence rolls it back. The original is never lost. Try background tasks : kick off a long job, then go do something else. Don't just stare at the screen. The longer you use it, the more accurately it tracks your preferences. That is the biggest difference between it and a "general-purpose AI assistant" — it works for you and you alone , and it gets to know you better the more you use it. If you run into a problem or a bug while using it, or there is a capability you wish it had, please email us. Tell us three things clearly: what you entered, what you saw, and what you expected — and we will get back to you quickly.