Chat & AI

The chat panel is how you interact with Qualia. Describe what you want in natural language and Qualia writes code, runs it, and shows you the results.

​

Basics

Type a message in the chat input at the bottom of the panel. Qualia will:

1. Read your message and understand what you want
2. Look at the current state of your notebook for context
3. Write and execute code cells to accomplish your goal
4. Show results (plots, tables, text) inline in the notebook

The chat input uses your system spell checker to underline spelling mistakes as you type. You can inspect and edit the notebook code Qualia runs. Expand an activity group in chat to see its actions and results; click a completed group’s summary to collapse it. Qualia maintains conversation history, so you can refer back to earlier results, refine approaches, and build on prior work naturally. When a long reply scrolls your message out of view, that message stays pinned in a single line at the top of the panel. Select it to jump back to where you asked. When a shell command prints a lot, the chat shows its first and last parts with a note for the lines in between. The full output is saved to a log in the workspace’s hidden .bg-logs folder (on cloud compute, on the machine that ran the command), and Qualia reads that log instead of re-running the command. A log keeps only the first 256 MB of a command’s output; past that it stops growing, but the command keeps running. Logs older than a week are removed automatically. While Qualia waits for a cell, command, task, or another agent, a wait row shows what it is waiting for. Reopening the chat restores an active wait without restarting the work. A row marked still running means the agent’s wait ended before the work finished. You can Queue follow-up while the agent is working. It reads queued input at its next decision step, which can be before the current cell finishes. A tool call already in progress may need to return first. Queued messages stay queued if you stop the agent. During a connection interruption, messages in existing chats wait to send after reconnection.

​

Starting from a path

The new chat screen offers three starting points that ask for a path first and then hand it to an agent:

• Explore a dataset asks for a single data file, then profiles it and reports its grain, its quality problems, and the questions it can actually answer. It adapts to what the data is, so tabular, time series, text, image, geospatial, event-log, multi-table, and graph data each get their own checks.
• Teach Qualia your workflow asks for a file or a folder, then turns a process you already run into a reusable skill.
• Import knowledge accepts files, folders, links, or notes and turns past work into claims.

In Qualia Cloud, the picker browses your cloud workspace. Use Upload or drag files onto the picker to add them to the current folder. Incoming files with duplicate names are renamed automatically.

​

Execution mode

Choose the default execution mode for new agents in Settings > Agents:

• Notebook uses an attached notebook for interactive code, cells, and kernel-backed results.
• Script creates and runs standalone files in the workspace. It is suited to training, scraping, batch jobs, and other work that does not need notebook cells.

The setting applies when an agent is created, so later changes do not affect existing conversations. Agents created from that agent inherit its mode, keeping the whole delegated workflow in notebook mode or script mode.

​

Basic things you can ask Qualia to do

​

Explore and clean data

• “Load sales_data.csv and show me the first 10 rows”
• “What’s the distribution of the price column? Show me a histogram”
• “Remove duplicates and fill missing values in the age column with the median”
• “Give me summary statistics grouped by region”

​

Build visualizations

• “Plot a scatter matrix of all numeric columns”
• “Show me daily revenue as a time series with a 7-day moving average”
• “Create a heatmap of the correlation matrix”
• “Make a bar chart comparing accuracy across the three models”

​

Train and evaluate models

• “Train a gradient boosted classifier on this data”
• “Compare random forest vs. logistic regression with 5-fold cross-validation”
• “Show me feature importances and the ROC curve”
• “Tune the learning rate and max depth using grid search”

​

Work with files and the shell

• “Install scikit-learn and xgboost”
• “What files are in the data/ directory?”
• “Run the preprocessing script at scripts/clean.py”
• “Search the codebase for where the model is saved”

​

Research and look things up

• “Search the web for the latest scikit-learn docs on HistGradientBoosting”
• “What’s the standard approach for handling class imbalance?”

​

More autonomous exploration

For open-ended research, give Qualia broader goals and let it explore:

​

Dataset investigation

• “Explore this dataset thoroughly — check for anomalies, correlations, and anything interesting”
• “Find the best features for predicting churn and explain why they matter”
• “Investigate why the model performs poorly on certain segments”

​

Model development

• “Build the best classifier you can for this data — try multiple approaches and compare them”
• “This model is overfitting. Diagnose the problem and fix it”
• “Optimize this pipeline for inference speed without sacrificing too much accuracy”

​

Research questions

• “What’s causing the spike in errors after March 2024? Investigate and report back”
• “Find a good forecasting approach for this time series data”
• “Review this codebase and suggest performance improvements”

These prompts work best with higher curiosity, independence, and parallelism settings, allowing Qualia to explore multiple angles, delegate suitable work, and make decisions without asking at every step.

​

Answering questions

Choose a single-choice option to send it immediately. Options appear as rows with inset horizontal dividers and rounded hover and selection backgrounds. The hovered row uses a stronger contrasting fill to make the active option clear. The divider beneath an option disappears while its blue selection background is visible and returns when that background clears. Hovering the selected option preserves its blue fill and hidden divider. The dividers immediately above and below the hovered option disappear; the remaining dividers stay visible and row spacing stays fixed. The hover background and Enter icon move together using the Files explorer’s 150 ms animation; reduced-motion settings disable the movement. Hovering over an option reveals an Enter icon. To write a custom answer, use the main message box below the question options, then press Enter or Send answer in its footer. Shift+Enter adds a new line. While a question is active, this box answers it instead of queuing a follow-up; there is no second free-response field. Draft answers are kept separately as you browse unanswered questions. Your ordinary message draft returns when the questions finish. The question and answer field share one outline with a straight divider. The answer box says “Your answer…”, and its footer preserves the attachment, model, model options, research-style, token-usage, and stop controls alongside answer submission while questions are active. Controls keep their usual availability during generation. “Use best judgment” uses quieter text than the other options. For multi-select questions, choose your answers, optionally add a custom answer in the main message box, and press Submit or Next in the composer footer. Questions with similar heights share a stable question area. Much shorter or taller questions resize gently to avoid large blank areas; longer content scrolls. Questions use a gentle fade rather than sliding; reduced-motion settings disable the fade. The card stays expanded while your answer sends. If sending fails, your draft remains available to retry. The top row shows the question count, such as 1 of 3, between previous and next arrows. Use the arrows to browse unanswered questions without submitting an answer. The question appears below and uses the full width. You have 20 seconds before the countdown starts and appears in a subtle pill at the upper right. Use the pause button to take more time, then the play button to resume. Typing a custom answer pauses it automatically for everyone viewing the question. Screen readers announce the remaining time when you focus the timer control. Collapsing the card does not pause the timer. When it expires, the agent uses its best judgment for unanswered questions. When Qualia asks a question, choose an option in the card above the chat box. To write your own answer, type it in the main message box. Multi-select questions let you choose several options before submitting. Choose Use best judgment to let Qualia decide.

​

Attaching context

You can attach specific context to your messages so Qualia knows exactly what you’re referring to:

• Cells: Select one or more notebook cells to include as context. Qualia sees both the code and its output.
• Files: Attach files and images from your filesystem for Qualia to read.
• Table selections: Select rows or columns from a data table and attach the selection.
• Highlighted text: Select text in a file, a notebook cell or its text output, or a terminal. Choose Add to chat in the selection toolbar, or press Cmd+L on macOS or Ctrl+L on Windows and Linux. A chip names the source; click it to reopen that source. Long pasted text can also appear as a chip. Qualia doesn’t send your current selection automatically, so add it when it matters.
• Earlier messages: Select any text in the chat — a passage of Qualia’s answer, or one of your own messages — and choose Reply from the button that appears. The quoted passage shows above the composer and stays attached to your message so Qualia knows exactly which part you mean. Notebook cells work the same way through the cell’s Chat about this cell action.
• Writeup passages: Select text and click the Qualia icon in the floating menu. Write your request in the comment on the right, then click Comment. Qualia replies in that thread with suggested edits you can accept or reject. Type @ to choose another agent by name. See Comments and suggestions.
• Images and regions: Hold Cmd on macOS or Ctrl on Windows and Linux while dragging directly on a plot output or chat image to select a region and attach just that part. When no region is selected, click once to focus the image, then click it again while focused to open the image viewer. When a region is selected, clicking the image clears it; use the expand button to open the viewer directly.

This is useful when you want Qualia to focus on a specific piece of your work rather than the whole notebook. Attachments can also be sent on their own: with a file attached and no message typed, sending will ask Qualia to look at the attachment. Files you attach appear inside the chat box while you compose. Remove one with its × button, or press Backspace at the start of the message to remove the most recent one. Attachments remain visible with your message after sending. Images show as previews — click once to focus a preview, then click again while focused to open it in the image viewer. Other files show as named chips; click a chip to download a copy of the file.

​

Model selection

Choose which AI model powers your conversation:

1. Click the model name at the bottom of the chat box, or the robot icon when the chat panel is narrow.
2. Choose a Frontier model, or select All models to browse the full catalog.
3. Use the search box to find any available model, including models outside the initial Frontier list.
4. To change which model new chats start with, hover a model and choose Set default. The current default is marked Default.

Auto is the simplest choice: it uses the best model for the task rather than making you name one, so it keeps working as models change. Use the options menu beside the model picker to adjust its reasoning or thinking level. In narrow panels, these settings appear inside the model menu. Available controls depend on the model; some offer effort levels without an off option. The catalog shows models available to your account. Some models also offer Fast mode. Fast mode prioritizes lower latency and uses premium provider pricing. With a bring-your-own OpenAI key, priority usage is billed directly by OpenAI. The provider may fall back to standard processing when a request is not eligible for accelerated processing, such as a long-context request.

​

Context window

The token usage indicator shows how much of the model’s context window you’re using. If you hit the limit, Qualia automatically summarizes (compacts) older history and continues, though the agent may lose detail from earlier in the conversation. To compact sooner, type /compact, optionally followed by what to keep, for example /compact keep the model comparison results. If the agent is running, compaction waits for its next step. Models also accept a limited number of images per request. In a chat with many images, Qualia omits the oldest images first and adds a notice explaining how to reopen them. Your chat history keeps every image. If you still reach the image limit, start a new chat.

​

Account usage limits

If your account reaches its weekly usage limit, Qualia stops the current request and shows a usage-limit message in chat. Use View pricing from that message to open Quadrillion pricing and choose a plan with higher limits.

​

When a model declines a request

If a model refuses a request for safety reasons, Qualia ends the response and names the model. Switch to another model or start a new chat. Resending to the same model may be declined again.

​

Research style

Adjust how Qualia approaches your requests (configurable in chat or as defaults in Settings):

​

Curiosity

How much Qualia explores beyond your immediate request:

• Low: Focused on exactly what you asked
• Medium: Explores related areas when relevant
• High: Actively investigates tangential questions

​

Independence

How much Qualia assumes versus asks:

• Low: Asks for confirmation frequently
• Medium: Makes routine decisions, asks about significant ones
• High: Assumes most things, asks only about major choices
• Auto ∞: Runs autonomously without confirmation dialogs

​

Parallelism

How much Qualia splits work across agents:

• Low: Usually keeps work in one agent
• Medium: Delegates independent work when useful
• High: Proactively splits suitable work across multiple agents

​

Auto-approve commands

Turn on Auto-approve commands to run shell commands and scripts without a confirmation card while keeping your chosen Independence level, so Qualia can still ask you questions. Library installs, MCP tools set to “Ask each time”, and catastrophic deletes still ask. Agents started from the chat inherit the setting.

​

Command approvals

The arrow beside Run in a command approval offers Always allow & run [command or tool] when there is one choice. With multiple command prefixes, the menu shows Always allow and run:, an All (N) checkbox showing the number of listed commands, and indented command checkboxes. Select the prefixes to always allow, then click Run. All selects or clears every prefix and shows a partial-selection state when only some are selected. Run and Skip are disabled while selected permissions save. If saving fails, an error appears and the command remains pending.

​

Autonomous Mode

Select Auto ∞ under Independence to run without confirmation dialogs. Qualia also stops asking you questions. Curiosity and Parallelism are dimmed while Auto is selected, and Auto-approve commands is shown on because Auto already covers it — see Autonomous Mode.

​

Code autocomplete

As you type in code cells, Qualia offers AI-powered autocomplete suggestions. These appear inline and you can accept them with Tab. Turn them on by setting the code completion mode to LM in Settings > Editor. Autocomplete uses the notebook context — nearby cells and their outputs — to make relevant suggestions. It requires signing in to Qualia and is not available on BYOK deployments.

​

Copying a message

Hover over a message and select Copy message beneath it. You can also reach the button with Tab; on touch screens it stays visible. Qualia’s copy button appears after the reply finishes. Replies copy both clean rich text and the original Markdown: formatted editors can keep headings, emphasis, lists, links, and code, while plain-text editors receive Markdown. The destination controls which formatting it accepts; tables and equations may look different in apps such as Slack. Human messages copy as plain text. Attachments, suggested replies, and read receipts are not included.

​

Fork a chat

Forking copies a chat into a new tab so you can branch off without touching the original.

• Whole chat: … menu > Fork chat.
• Up to a reply: hover over a reply and choose Fork in new chat.

Auto-approve isn’t copied, so commands in the fork ask before running.

​

Exporting a chat

Open the … (Chat actions) menu at the right end of the chat tab strip, or right-click a chat tab, and choose Export to save the conversation as a Markdown (.md) file. The export contains the visible transcript — your messages and Qualia’s replies — so you can archive a conversation or share it outside the app. Export works in both the desktop app and Cloud. Cloud exports include participants’ names beside their messages. Hidden prompts, reasoning, tool results, and internal agent messages are excluded.

​

Customization

Qualia’s behavior can be customized with:

• Rules: Persistent instructions that apply to every conversation (e.g., “Always use polars instead of pandas”)
• Skills: Reusable workflows for specific tasks (e.g., “How to deploy to production”)

See the individual pages for details on creating and using these.

​

Collaborative chats

In a shared cloud workspace, click the small person/people Chat access button beside the chat tabs. Under Who can message this agent?, the creator can choose Only you or Everyone in this workspace. The selected option shows the current access; workspace members can always view the chat. Messages show each sender’s identity, with typing and read indicators. The agent keeps its creator’s permissions and usage. See working together in a chat. If you are viewing someone else’s chat, Chat access shows Only the creator or Everyone in this workspace, and only the creator can change it.