Help & Guides

Chat is the command center where your work with Alkimi comes to life. It is the interface where you interact with your agents, but it is more than just a messaging window; it is a collaborative workspace. Here, you can seamlessly weave together your proprietary knowledge, real-time web data, and uploaded files into a unified conversation.

Designed for productivity, the Chat interface gives you powerful tools to manage your workflow. You can easily switch between different specialized agents, organize your history with pins and archives, and use advanced features like temporary chats for privacy. Whether you're brainstorming strategy, analyzing a document, or running a complex query, Chat provides the structure and tools you need to get the most out of your AI workforce.

Managing Conversations

The chat sidebar is designed to help you keep your conversations organized. It features a tabbed interface to filter your chats and a context menu for quick actions.

History and Archived

• History: Your main inbox for non-archived chats. Pinned conversations stay at the top in a Pinned section; everything else is grouped by date (Today, Yesterday, Previous 7 days, Previous 30 days, and Older).
• Archived: Conversations you are done with but want to keep. Archived chats are hidden from History and use the same date groups in this view.
• Switching tabs clears the list and shows a spinner until the new results load. Switching again before the previous request finishes cancels it and loads the newly selected tab.

Chat Actions

Inside an open chat, the Actions menu () in the header offers the same Share, Rename, Pin, Archive and Delete actions plus Print, which lays the conversation out for paper or PDF. You can also click the chat title in the header to rename it in place.

Message Actions

Hovering a message reveals its action bar. Every message offers Copy message (badges and charts survive the copy, see Re-using files) and, when the agent enables quotes, Quote message to reply to a passage. Your own messages have Edit message: change the text or attachments and Send as new branch; the original message and its replies are kept. Agent replies have Regenerate response, which also creates a branch, and Share from this message. Where a message has several branches, a small navigator under it lets you step between them. Less common actions sit in the More message actions menu.

Temporary Chats

For quick, one-off conversations that you don't need to save, you can create a temporary chat by clicking the dashed message icon () next to the "New Chat" button, or by hovering an agent in the New Chat dropdown and clicking the temporary-chat button that appears on its row. Guests always get temporary chats. These chats are never shown in your History and cannot be reopened once you leave, so you are asked to confirm before navigating away. Temporary chats cannot be pinned, archived or shared.

Searching Chats

The Composer

The message box at the bottom of the chat does more than accept text. Its controls let you choose which agent and model answer, pull tools and knowledge collections into the message, and open the chart builder or the image editor.

Press Enter to send and Shift+Enter for a new line. A message can hold 5,000 characters by default (an agent manager can raise the limit up to 100,000); a live counter appears as you approach the limit and Send is disabled while you are over it. The Refine button left of Send (or Tab once you have typed something) opens the prompt helper, which asks a few clarifying questions and rewrites your draft into a stronger prompt that you can apply or dismiss.

Agent and Model Picker

The pill at the bottom-right of the composer names the active agent and, when the agent allows it, shows the language model and image model in use as their brand marks. Hover a mark for a mini card with the model's name, tier, Alkimi Intelligence score, context window, data policy and estimated cost; hover the agent name for the same summary the Agents page shows. Click the agent or a mark to open the picker on that section.

• Agent: Lists the agents you can chat with. On a new chat, picking another agent starts the conversation with it and keeps your typed message and attachments. On an existing chat the list is locked (an agent is chosen when a chat starts); the footer offers New chat instead.
• Language: Shown when the agent allows model switching. It starts with a short curated list: the Alkimi tier routers (Alkimi Auto, Basic, Standard and Premium, each labelled "Alkimi router") plus the agent's default model. More models opens every model the agent allows. Models over the agent's spend limit, or lacking an input the agent requires, are shown locked with the reason.
• Image: Shown when the agent allows image generation and image-model switching.

Every tab is type-to-search (name, provider, tier, or a capability such as pdf, reasoning or zdr). Hovering a row (or pressing Right) opens a detail card that compares the model with the agent default. A pinned language or image model applies to this chat only and is saved with the chat; a Reset to default row appears while a model is pinned. The footer always shows the per-message credit range for the active agent.

Tools Menu and @Mentions

Two controls insert tools and knowledge collections into your message:

• Type @ to open the mention list at the caret. It offers the agent's built-in tools (web, and math when the math keyboard is enabled), every collection the agent lets you mention, and a Files in this chat group (see Re-using files). Keep typing to filter; matching is anywhere in the name, so post finds "OCR Post-Processing".
• Tools button (sliders icon, left of the composer) opens the same set in a menu. Its first level lists the built-in tools plus a Collections row to drill into. Below a divider are actions that open a flow instead of inserting text: Sketch an image opens the image editor on a blank pad (requires image generation and file uploads on the agent) and Insert a chart opens the chart builder.

The paperclip (attach) button and its menu are only for adding files: upload from your device, or pick from Collections and connected cloud drives. Opening one composer menu closes the others.

Suggested Questions

To help guide the conversation, the agent can provide a list of suggested questions. These AI-generated prompts may appear at the start of a new chat or after an agent's response to help you explore the topic further. You can customize how these suggestions are displayed from your Account Preferences.

Sharing Conversations

You can share conversations with others via a secure, read-only link. This is useful for collaboration, getting feedback, or simply sharing an interesting interaction with an agent. When you share a conversation, you are creating a static snapshot of the chat up to a certain point. The shared link will not update with new messages.

How to Share

There are two ways to share a conversation:

• Share Current Chat: Open the Actions menu () in the chat header, or the chat's row menu in the sidebar, and choose Share ().
• Share a Snapshot: To share a conversation only up to a specific point, hover an agent reply and click Share from this message () in its action bar; it is also listed in the More message actions menu. The button shows how many shares already start at that message. This creates a link that includes that message and all previous messages.

Both methods will open the sharing dialog, which provides several options to control how your chat is shared.

Sharing Options

• Visibility: You can control who can view the shared link.
  • External: Anyone with the link can view the conversation.
  • Organization: Only logged-in members of your organization can view the conversation.
• Scope: Choose how much of the conversation the link includes. Sharing from the header offers only Current chat. Sharing from a message offers Up to this message, This exchange (the reply together with the prompt that produced it) and This message only.
• PII Safety: The system automatically checks for personally identifiable information (PII) and warns you if potentially sensitive data (like names or emails) is detected in the chat. Turn on Remove sensitive info to scrub detected PII with placeholder tags. Additionally, to protect your privacy, your name as the sender is automatically hidden on messages when sharing with "External" visibility, though it remains visible to colleagues on "Organization" shares.
• Branches & Citations: Include all branches also shares existing regenerations and edits rather than only the path you are viewing, and Include citations shows inline knowledge references in the shared view.
• Custom Label: You can give your shared link an optional, memorable label to make it easier to identify later.

Managing Your Shared Links

The sharing dialog also lists the links that already exist for the chat, each with its view count, and lets you Copy link, rename, delete one link or Delete all of them. You can manage all of your shared links from the My Shares page in your account settings. This page provides a centralized view of every share link you've created.

• View & Copy: See a list of your shares, including their custom names and the chats they belong to. You can easily copy the link to your clipboard.
• Delete: You can revoke access to a shared chat at any time by deleting its link. Once deleted, the link will no longer be accessible. You can also delete all of your shares at once.

What the Recipient Sees

When someone opens your shared link, they will see a clean, read-only version of the conversation. The page includes the chat title, the agent's name, and the date it was shared. If the share is an "Organization" type link and the viewer is logged in, your name will also be displayed as the person who shared it. A notice will be shown if PII was removed from the content. Logged-in viewers of an "Organization" share can also leave comments on the shared messages; comments are not available on "External" shares.

Uploading Files & Images

If an agent has file uploads enabled, you can attach files and images to your messages. This allows the agent to analyze the content of your documents, answer questions about them, or use the information within them to perform tasks.

How to Upload

To upload a file, click the paperclip icon () in the chat input bar, drag files onto the composer, or paste them. You can upload multiple files at once, up to 50 files and 20 MB in total per message (the composer warns you when a selection goes over either limit). Supported file types include:

• Images: PNG, JPG, GIF, WEBP, and SVG. An SVG is sent to the model as its markup (so any text model can read and edit it) together with a rendered preview when the model accepts images.
• Documents: PDF, and Office documents (Word, PowerPoint, OpenDocument; DOC/DOCX, PPT/PPTX, ODT/ODP), which are converted to PDF on upload.
• Spreadsheets: CSV and Excel/OpenDocument workbooks (XLS/XLSX/ODS), which are converted to CSV on upload and preview in a spreadsheet grid.
• Text & Code: TXT, Markdown, LaTeX (.tex, .latex, .ltx, .sty, .cls), JSON, XML, HTML, YAML, and common source-code files.
• QTI packages: A .zip QTI package can be attached when the agent has QTI import enabled.

Attached files appear as small badges inside your message text at the point where you added them (the caret, or the drop point when you drag a file in). While a file uploads its badge shows a progress fill; a failed upload shows a warning badge with the reason in its tooltip. You can drag a badge to another spot in the text, remove it with its x (or delete it like text), and undo the removal with Ctrl+Z. Hovering an image badge floats a thumbnail; clicking any badge opens a preview.

Viewing and Managing Files

Re-using Files Already in the Chat

A file that is already in the conversation never needs to be uploaded again:

• @ menu: Typing @ lists a Files in this chat group with the files on your draft and the most recent earlier files. Picking one inserts a reference badge at the caret.
• Copy and paste: Copying a sent message keeps its badges; pasting into the composer turns each one back into a badge for the same file.
• Paste from another chat: Files are never linked across chats. A badge pasted from a different conversation is copied into this chat as a new upload and validated for this agent like a file from disk. If the current agent cannot accept it, the badge turns red with the reason and Send stays disabled until you remove it or switch to an agent that accepts it.
• Drag an image out of a reply: Generated images, knowledge figures and rendered SVG code blocks can be dragged straight into the composer text. An image that is already a stored file becomes a reference badge; an SVG code block is attached as a new .svg file.

The model receives each file's content once; later references arrive as a short label and do not count against the file capacity below. The same rule drives the pages/images counter, so re-attaching a file you already sent never uses more capacity.

Understanding Upload Capacity

Each agent with file uploads enabled has a maximum upload capacity, which is determined by its Message Spend Limit. This capacity is measured in "pages" (for documents and images). As you attach files, a progress ring around the attachment button will show you how much of the available capacity you are using. This ensures that the total size of uploaded files does not exceed what the agent can process in a single message.

Image Editor

The image editor is a full-screen dialog for sketching a new image or marking up an existing one before sending it to the agent. It opens from Sketch an image in the composer's Tools menu (blank pad), from Edit on a generated image in the thread, or from the lightbox on an image you uploaded. Sketching and every generative action require the agent to have image generation and file uploads enabled; instant tone and frame edits can be attached or downloaded without involving the agent at all.

Four tabs group the work by intent. Undo, redo and hold-to-compare apply across tabs, and the footer's primary button follows the active tab. Keyboard shortcuts: hold Space to compare with the original, press 1-4 to switch tabs, E to leave the preview and start editing, Ctrl/Cmd+Z to undo and Ctrl/Cmd+Y (or Ctrl/Cmd+Shift+Z) to redo, and Escape to close.

• Pins: Drop up to five numbered pins on the image and type what should change at each spot. Draw and Text add strokes and labels as extra guides; the Eraser removes whole marks. On a blank sketch pad the same tools sketch the scene, joined by Fill (a paint bucket bounded by your lines). You can also upload or drop an image to draw over. Write a short description (up to 500 characters), then Generate sends the sketch to the agent or Attach to message stages it as a PNG with your description prefilled.
• Remove: Paint a mask and ask the agent to Remove the area, Replace it with something you name, or Cut out the subject in one click.
• Frame: Crop with aspect presets, rotate, flip and straighten. When the frame extends past the image, Expand with AI asks the agent to fill the new margin.
• Colour: Instant looks and light sliders (brightness, contrast, saturation, warmth) preview live and can be attached or downloaded without a model call. Below them you can stage an AI restyle or an Upscale (offered only when the chat's image model can output a larger resolution than the image has).

Generative actions upload the working image (with frame, tone and marks baked in) and message the agent; they are billed like any image generation. Pin instructions travel in the message text.

Charts & Math

Charts render as interactive graphics directly in the conversation. An agent can produce them in its replies (ask it to plot a function or chart some values) and you can build your own with the chart builder. Charts support zooming, hover tooltips, touch gestures and a fullscreen view, and adapt their colours to light and dark mode.

Chart Types

• Basics: line, area, bar (including range bars for Gantt-style timelines), scatter, bubble, waterfall.
• Parts of a whole: pie, doughnut, radar, polar area.
• Distributions & finance: histogram, box, violin, candlestick.
• Fields & 3D: heatmap, contour, 3D surface, parametric surface, vector field (arrows and streamlines). Surfaces can be rotated by dragging.
• Functions: y = f(x) and x = f(y) curves (example: x^2), parametric curves in t (example: x = cos(t), y = sin(t)) and polar curves in theta (example: r = 2 sin(4 theta)). Discontinuities are drawn with open and filled endpoint dots, and asymptotes can be marked with dashed guides.
• Relationships: sankey, graph (vertices and edges, with an arrangement menu and draggable vertices in fullscreen), parallel coordinates.
• Clips: any expression chart can be animated over a time variable; a clip plays with its own controls and can be downloaded as a video at 1080p, 1440p or 4K.

Chart Builder

Insert a chart in the composer's Tools menu opens the chart builder. Pick a Chart type from the library above and the rest of the form appears: Title, X axis label, Y axis label and an optional fixed Y axis range; the data editor that matches the type (Data rows and series you can paste from a spreadsheet, per-record columns, a grid of z values, or Curves/Expression fields with ranges); Options (legend, stacking, smoothing, colour ramp, contour levels, mesh, envelopes and asymptote guides, and so on) with an Animate checkbox for clips. The Preview updates as you type, and the footer button reads Insert chart (or Update chart when editing an existing chip).

Inserting places the chart in your message as a chip that renders as an interactive chart once sent; click the chip to edit it. Charts already in the conversation (yours or the agent's) have an Open in chart builder button in their header that opens a copy for you to tweak and re-insert; the original message is not changed. Copy chart puts a chart on the clipboard: pasting into the composer adds it as an editable chip, while rich targets such as documents receive it as an image. Copying a whole message, or a selection that includes a chart, works the same way. The header also offers Fullscreen, Copy as TSV (the plotted values as tab-separated text for spreadsheets), Copy as Image and View as table, which swaps the graphic for its data table in place.

Supported Functions & Arguments

Expression charts accept a wide range of mathematical functions. Most functions take a single argument (x), but some require additional parameters. Every standard JavaScript Math function (including pow, hypot, atan2 and random) is also available.

Supported Operators

Note: Both ^ and ** mean exponentiation and bind tighter than unary minus, so -x^2 is $-(x^2)$ as in mathematics and 2^3^2 is $2^{9}$. Bitwise operators are not supported. Comparisons return 1 for true and 0 for false.

Using Tools

When you're chatting with an agent that has tools enabled, you can invoke them to bring external information into the conversation.

Using @web

If Web Search is enabled for the agent, type @web in the current message and accept the autocomplete suggestion to attach a web badge. The badge forces a search for that message. Adding sources to the badge also restricts the search to those sources.

• Disabled: The agent cannot search the web, and @web is unavailable.
• Explicit: The agent searches only when the current message contains a valid @web badge.
• Implicit: The agent may decide when a search is useful. A valid @web badge in the current message always forces a search, and a scoped badge both forces and restricts it.

Choose and Edit Sources

The picker presents one flat source list, grouped under headings. Each row is a single domain that you toggle individually; the headings are not selectable as a whole:

• Recent: Your last 10 actually used domains across the account.
• Your organization: DNS-verified and manually added organization domains.
• Research and scholarship: Academic indexes, open-access journals, and scholarly repositories.
• Teaching resources: Open courseware, textbooks, and educator-created learning materials.
• Reference and archive: General reference and archived content.
• Custom domain or URL: Enter a public domain or paste a URL. A pasted URL is normalized to its hostname. Type , (comma) to commit the typed domain and keep adding more.

A domain includes all of its subdomains. Where the picker allows them, .edu and .gov act as suffix filters. Choose 1–3 websites for a focused search; one badge can contain at most 5. Organization sources are suggestions and do not restrict search unless the agent has a Restricted source policy.

Multiple Badges and Recents

Multiple badges keep independent source sets. For example: [web | arxiv.org] Compare current research with [web | ed.gov, .gov] federal guidance. Equivalent sets are deduplicated regardless of domain order, and the query helper can select only among the supplied sets.

Badge labels adapt to available width by showing domains followed by +N. A solid divider separates web from the scope, and the tooltip lists every source vertically. Use the quick X to remove one recent domain. Manage or clear the account-wide list in Account Preferences.

Search activity runs in order: Search, then Browse for selected pages. Scoped Google links include the enforced site: filters. Search and browse operations are billed when executed; see the agent's current Tools cost display and the Billing & Pricing documentation for current rates.

Tips for Effective Web Searching

Source badges are enforced boundaries. Natural-language requests to prioritize or exclude sources are guidance to the model and are not guaranteed restrictions. Use a scoped badge whenever the source boundary matters.

1. Be Explicit with Multiple Queries: If you need to compare or gather information on multiple topics, guide the agent to perform separate searches. This yields more focused results than a single, broad query.
   Example: Instead of asking Compare Google's Gemini Pro with Anthropic's Claude Sonnet, a more effective prompt would be:
   Please perform web searches for the technical specifications of Google's Gemini Pro and Anthropic's Claude Sonnet, then summarize the key differences.
2. Enforce a Source Scope with Badges: Select domains when results must stay within a known set.
   • First-party example: Select the official domain, then ask your question.
     Example: [web | nextjs.org] What are the latest features in Next.js?
   • Category example: Select the domains listed under a built-in heading such as Research and scholarship. When your selection matches that heading's full set exactly, the badge shows the short group name instead of the domains.
     Example: [web | Research] Summarize recent work on retrieval-augmented generation.
3. Use Natural Language for Preferences: Phrases such as prioritize official sources, favor independent reviews, or avoid pinterest.com can guide selection inside the enforced scope, but the model may not follow them perfectly.
4. Use Suffix Filters When Available: Select .edu or .gov when the result set must remain on that suffix.
   Example: [web | .edu, .gov] Find guidance on sustainable farming practices.

Understanding Web Search Limits

To ensure performance and relevance, the web search tool operates within a set of technical limits. It's helpful to be aware of these as you use the feature.

• Search Queries per Message: The agent will generate a maximum of 5 distinct search queries based on a single message from you.
• Websites Visited per Message: From the search results, the agent will select and "visit" (read the content of) up to 10 websites in total for a single message, drawn from a shared pool of at most 3 pages per search that ran (so two searches allow up to 6 pages, four or more allow the full 10).
• Budget: Both caps are lowered further when the agent's Message Spend Limit leaves fewer credits for searching and browsing, since each search and each page browse is charged before the model runs.
• Links in your message: If URL browsing is enabled for the agent, up to 10 links pasted into your message (plus a few links from recent messages in the chat) are fetched directly, ahead of any search, and each fetch counts against the same budget.
• Total Content Limit: The agent also has a reading limit tied to its context window. Pages that do not fit the remaining context budget are dropped from the prompt silently: they still appear as "Browsed" in the status list, but the model does not see their text. This mechanism prevents the agent from being overloaded with information that exceeds its context capacity.

Monitoring the Process: Understanding Statuses

When an agent uses a tool like Web Search, you'll see a series of status updates in the chat window. These are designed to give you real-time insight into the agent's actions.

• Searching: The agent is actively performing a web search for the given query.
• Searched: The agent has completed the web search for the query.
• Browsing: The agent is currently reading and processing the content of the linked URL.
• Browsed: The agent has finished reading the content of the linked URL.
• Skipped search: A scoped search query was not run because none of its sources were permitted, for example when the agent's Restricted source policy allows none of the domains in the badge. Browsed pages are never marked skipped; see the Total Content Limit for what happens when a page does not fit.
• Failed: The agent was unable to read the content of the URL. This can happen for various reasons, such as the website being down, access being forbidden, or the page containing no readable text.

Each status includes a clickable link that allows you to view the relevant search results or webpage in a new tab, giving you the same context the agent is working with.

How the Agent Selects URLs to Read

After performing a search, the agent uses an AI model to analyze all the resulting links and selects only the most promising ones to read. You will only see status updates for the URLs the agent decides to browse. Links that are deemed irrelevant by the model are discarded and will not appear in the agent's action steps.

Message Costs

If you can view the agent's configuration, each agent reply carries a small credit badge. While a reply streams the badge already shows the cost the message will settle at, so the figure does not change after a reload. Hover the badge (or long-press it on touch devices) for a breakdown:

• Message generation: what the model read (instructions, history, knowledge, web pages, attachments) and what it wrote back, split into Input, Reasoning and Output sub-rows.
• Retrieval & context: the searches, URL selections and page browses that ran for this message, in order, each with its charge.
• Image generation, References and Suggestions: separate rows for any images the reply generated and for the helper calls that produced citations and suggested questions.
• Composition bar (agent managers): how the message's prompt window was filled, with an Open room segment for the space left unused. It is the per-message counterpart of the Where Credits Go card on the agent's Statistics page.

When a reply is cut off at its output limit, agent managers also see a warning badge that says which limit applied: the agent's Message Spend Limit (raise the slider on the agent's Model page) or the model's own output limit, in which case a higher spend limit would not help and the badge suggests switching model or asking for the answer in parts. The same badge appears when a reply used more than 75% of its context window; its tooltip then shows the Context Usage percentage and recommends raising the spend limit.

Display Preferences

How the thread is rendered is a personal setting that follows your account across devices. Under Account Preferences you can choose between the Bubbles style (the default: every turn in a tinted message box) and the Flat style (replies read directly on the page under the agent's name). Flat adds a Brand-tinted blocks switch that washes code, math, callout and table blocks with the agent's brand colour. You can also set the Reading width (Comfortable, Wide or Full width), the Message density (Comfortable or Compact), Show avatars (off by default) and Always show timestamps (on by default; turn it off to reveal times only on hover). Embedded agents and public share links always use Bubbles.