AI Assistant
SiteCAD includes a built-in AI assistant that understands your site, your data, and your tools. It can query your scenario, modify entities, write scripts, design solids, and help you work faster through natural conversation.
Conversation Basics
The AI assistant lives in the Agent panel on the right side of the editor. Open it, type a message, and the assistant responds with text, data, or actions — whatever fits your request.
Context-Aware Conversations
The assistant automatically knows about your current scenario — its entities, terrain, properties, and structure. You don’t need to describe your site from scratch. Ask questions like “How many boundaries do I have?” or “What’s the total area of all parcels?” and it queries your data directly.
Four Contexts
The assistant adapts its tools based on what you’re doing:
- Scene context — The default. Query and modify your entire scenario: entities, terrain, properties, and data.
- Solid context — Active when you’re editing a parametric solid. The assistant can read and write blueprints, parts, and parameters.
- Operation context — Active when you’re editing an operation script. The assistant can update code, parameters, and run the script.
- Drawing context — Active when a drawing sheet is open. The assistant can lay out views, annotate them, fill in the title block, apply templates and export the sheet. Every sheet shares this one conversation.
Persistent Conversations
Conversations are saved locally per editor context. Close the panel or navigate away — your conversation history is preserved when you return. Start a new conversation anytime to reset the context.
Interactive Results
When the assistant returns SQL queries or code, the results are interactive. You can copy code blocks, run SQL directly in the console, and expand or collapse tool results to focus on what matters.
Available Tools
The assistant has different tools available depending on your current editing context. It chooses which tools to use based on your request — you just describe what you want.
Scene Tools
Available in the main editor view. These tools work with your entire scenario.
| Query | Run read-only SQL SELECT queries against your scenario database. Browse entities, inspect properties, or analyze data without changing anything. |
|---|---|
| Command | Execute SQL mutations (INSERT, UPDATE, DELETE) to modify your scenario. May require approval depending on your settings. |
| Run Script | Execute JavaScript with access to geodetic transforms, coordinate conversions, and file operations for complex multi-step tasks. It is also how the assistant runs one of your saved operations — your own or one installed from a library — by name, with the parameters you give, and can feed one operation’s result into the next. |
| Get Library | Fetch reference documentation on solids, operations, toolkit APIs, and parameters so the assistant can give accurate answers. |
| Web Search | Search the web for external information when enabled. Useful for looking up material properties, regulations, or reference data. |
Guiding Your View
In the scene, the assistant can point things out and move your view as well as your data. These change only what you are looking at, not your scenario, so they never ask for approval, even with Require Approval for Mutations on.
| Select | Selects the entities it is talking about, exactly as if you had clicked them, and can fly the camera so they are all in view. Anything it names that no longer exists is skipped, and if none of it exists your own selection is left as it was. Back returns to the view and selection from before it flew. |
|---|---|
| Isolate for editing | Makes an entity the active editing scope, the same as choosing it in the scene graph’s Active editing scope picker, or returns to the whole scene. This is separate from selecting: an entity can be selected without being isolated, and the other way round. Back returns to the previous scope. |
| Back / Forward | Steps your view history, as the toolbar’s Back and Forward buttons do, restoring the scope, selection and camera of that point: “take me back to where I was.” |
| Home, 2D and 3D | Flies to the site overview as the Home key does, or switches between the 3D globe and the flat 2D map, as the Camera menu’s View row does. Neither is a step Back undoes. |
| Standard views and projection | Turns to a top, north, south, east, west or isometric view, snaps the heading to the nearest compass point, or switches between perspective and orthographic, as the Camera menu’s Look and Projection rows do. A standard view is one Back step; the snap and the projection switch are not. |
| Show a section | Shows a drawing’s section or elevation in the scene, squared level and orthographic to its cut line, as the view inspector’s Show in scene does: “show me the north elevation of the house.” One Back step. |
| Place the camera | Reads where the camera is, or puts it at a position and heading, or aims it at a point, as the Camera menu’s Position… dialog does. One Back step. |
In the conversation
Each of these leaves a card in the conversation. A Selection card lists what it selected as links: click one, or Select again, to select and frame them again; a link to something since deleted says so instead. An Editing scope card has Return to this scope. View history and camera cards have no button, because the toolbar and Camera menu are on screen. If you have unsaved changes in the solid editor, the assistant will not change the scope or step the history; it tells you instead, so nothing is discarded.
Solid Editing Tools
Available when editing a parametric solid. The assistant can see and modify your solid’s scripts and parameters.
| Get Solid Script | Retrieve all blueprints, parts, and parameters for the current solid. |
|---|---|
| Upsert Script | Create or update a blueprint or part script. Returns execution results so you can see the effect immediately. |
| Set Parameters | Update the parameter schema — add, rename, or change parameter types and defaults. |
| Save | Persist script changes to the database. |
| Remove Script | Delete a blueprint or part that is no longer needed. |
Operation Editing Tools
Available when editing an operation script. The assistant can update code, manage parameters, and run your operation to test it.
| Get Script | Retrieve the current operation script and its parameter definitions. |
|---|---|
| Set Script | Update the JavaScript code. May require approval depending on settings. |
| Set Parameters | Update parameter definitions for the operation. |
| Run Script | Execute the operation and return structured outputs for review. |
| Get Errors | Retrieve the last error from the operation editor for debugging. |
| Save | Persist operation changes to the database. |
Drawing Tools
Available while a drawing sheet is open. Drawings are part of the Studio plan. From the 3D scene, without a sheet open, the assistant can still create a sheet and open it for you.
| Add Views | Add a plan, section, elevation, profile or detail view, or cut a set of cross-sections along a trace, which can fill new sheets. |
|---|---|
| Configure Views | Change a view’s framing, style, scale or name, or remove a view. Removing a view may require approval depending on your settings. |
| Annotate | Add and edit dimensions, leaders, text labels, slope labels, volumes tables and spot elevations. |
| Sheet & Title Block | Set the page size, scale, name and notes, and fill in the title block. |
| Templates | List and apply the built-in templates or your own. |
| Export | Export the sheet as SVG or PDF. |
| Close | Close the sheet and return to where you opened it from. |
Skills: reference the assistant reads when it needs it
The assistant has reference guides, called skills, each on one area of SiteCAD: the scene, solids, operations, drawings, libraries, settings and more. It does not send them all with every message; it reads one only when your request needs it, which keeps each message small. A card headed "Skill:" in the conversation shows a guide it has just read. That is a normal step of its reply, not an error.
- While a guide loads, the status row above the message box reads "Reading a skill guide…". The Skill card that follows holds the full text the assistant read; expand it to see what it worked from, or copy it.
- Skills are part of the app, or of your site's data on this device. Loading one is not a download or a web search.
- Some skills also give the assistant extra tools. It can use them from the next step of the same reply, so reading a skill before acting is expected.
- A loaded skill stays loaded for the rest of the conversation; nothing unloads it. Its text goes with every later message, so each one adds to what later messages cost. A new conversation starts with none.
- No screen lists the built-in skills. Ask the assistant and it will name the ones it has.
- A site can carry skills of its own: rows of the scenario's Agent Skills table in the Data tab, added there or brought in by an installed library. They load the same way and change with the open scenario. To give a library you publish a skill of its own, see the ⓘ beside By table in its Publish dialog.
The ⓘ on a Skill card in the conversation opens this same text.
Attaching Files
The assistant can’t open files on your computer itself — it works only with the files you attach to a message. When it needs a file to continue, it asks you to attach one.
| Browse | Click the paperclip button beside the message box and choose one or several files. |
|---|---|
| Drop | Drag files from your computer anywhere onto the Agent panel. |
| Paste | Paste a screenshot or a copied file into the message box. If the clipboard also holds text — cells copied from a spreadsheet, for example — the text is pasted as usual instead. |
Files waiting to be sent show as chips above the message box; click a chip’s × to remove it. You can send files with or without a message, and Send waits until every file has been read. Each file can be up to 100 MB. Files you haven’t sent are not kept with an unsent draft: switching to another conversation drops them. Each sent message lists the files it carried.
What the Assistant Does with a File
- Images — An attached image is shown to the assistant with your message, so it can look at a screenshot straight away. Later in the conversation it opens the image again when it needs another look.
- Data files — The assistant reads any other file with its tools — the whole text, or a preview of a large file — and can import it into your scenario, including formats the Import dialog doesn’t recognize.
- GeoTIFFs — The assistant imports a GeoTIFF as terrain or as imagery, and asks which if you haven’t said.
- PDFs and Word documents — The assistant reads these page by page, zooms in and measures drawings; see Reading Documents, below.
For this session only
Attached files are not saved with your site — only their names are. The assistant can read a file for the rest of your working session. After you reload the page, or once a file has gone unused for a while, the conversation still lists it but the assistant can no longer open it, and asks you to attach it again.
Reading Documents
An attached PDF, image or Word document is something the assistant can read, not only import, in the way a person reads it: a shop drawing, a site plan, a specification. As soon as one is attached it knows the page count and sheet sizes, an image’s size in pixels, or for a Word document its length and how many headings and figures it has.
| Read the text | Reads the text of a PDF or Word document, whole or a part at a time, and finds where a word, label or detail number appears. A scanned page, or a CAD export whose text was drawn as outlines, has no text to read, so the assistant looks at it instead. |
|---|---|
| Look closer | Looks at a whole page, or zooms into a region of it until small dimension text is legible. It looks at attached images and at the pictures embedded in a Word document in the same way. |
| Measure linework | On a PDF exported from CAD, it reads the exact lines, arcs and circles in a region instead of estimating from the picture, applies the sheet’s scale, and checks the result against the dimensions written on the sheet. A written dimension wins; if the two disagree, the drawing may not be to scale. A scan has no linework, so anything measured from one is an estimate from the picture. |
| Ask a separate reader | For a question that needs many views or measurements, such as a detail’s dimensions, a schedule or a hole pattern, it can hand the question to a separate reader on the same model. A Document reading card in the conversation shows the question, the answer, the pages and regions the reader looked at, and how sure it was. The reader’s pictures are not resent on every later step of the reply. |
Views aren’t kept
A page the assistant looked at is not kept in the conversation, so on a later turn it looks again. One reply shows it at most 16 pictures, counting images sent with your message. If a drawing needs more, it summarizes what it has and continues on the next turn.
Supported Formats
- Images — PNG, JPEG, WebP and GIF (a GIF’s first frame).
- Word —
.docx. Tracked changes are read as accepted; comments and headers and footers are not read; pictures in EMF or WMF format, common when pasting from CAD, cannot be shown. The assistant says so if one matters.
For an old .doc file the assistant asks for a .docx or PDF
export; for a password-protected document, a copy without the password; and for a Word
document too large to open, a PDF export. PowerPoint and OpenDocument files, and TIFF
images other than GeoTIFFs (which import as terrain or imagery; see Attaching Files), are not read as documents, so export them to
PDF or an image first.
Approval Workflow
When the assistant wants to change your scenario, you can choose whether to review those changes before they’re applied.
When Approval Is Required
Turn on “Require approval for mutations” in settings and every data-modifying action shows a preview first.
How It Works
The assistant proposes a change (SQL mutation, script update, or parameter change).
A preview appears showing exactly what will happen — the SQL statement or code diff.
Click Approve to apply the change, or Reject to discard it.
The assistant sees your decision and continues accordingly.
Tip
Even without the approval setting, all changes are undoable. Press Ctrl+Z (or ⌘+Z) to reverse any action the assistant took.
Settings
Configure how the assistant behaves. These settings are stored locally in your browser.
| Web Access | web_access_enabled Allow the assistant to search the web for external information like material properties or regulations. |
|---|---|
| Auto-Execute Queries | auto_execute_queries Let read-only SQL queries run automatically without showing results inline first. |
| Require Approval for Mutations | require_approval_for_mutations Show an Approve/Reject prompt before any change is applied. Off by default (changes run immediately, like Operations, and are easy to undo); turn on for extra caution on shared data. |
| Site Prompt | site_prompt Custom instructions that are sent with every message. Use this to set context like "This is a 40-acre homestead in central Texas" so you don’t have to repeat it. |
Site Prompt example
Set a site prompt like “This is a 40-acre homestead in Hill Country, Texas. The north boundary follows a creek.” The assistant will use this context in every response without you repeating it.
Tips & Best Practices
Get more out of the assistant with these approaches.
Be specific about what you want
"Add a 10m × 5m concrete pad at control point CP-3" works better than "add something near that point."
Ask it to explain before changing
"What entities are within 50m of the boundary?" before "Move all entities 50m north."
Ask what lies inside a boundary
"Which trees are inside the cut-limit boundary?" or "How much fencing would the paddock boundary need?" It counts a hole as outside, and lists separately anything it cannot place.
Use it for data exploration
"How many control points have elevation above 500m?" or "Show me all boundaries with area greater than 1 hectare."
Let it write scripts for you
"Write an operation that calculates cut/fill volumes between the existing and proposed terrain."
Run operations you already have
"Run my Irrigation Schedule operation for zone 3 with a 20-minute duration." It finds the operation, runs it with those values, and shows the output here. With approval on, each run asks first.
Ask for format recommendations
Attach the .xml file your surveyor sent and ask: "What’s the best way to import this?"
Work from a drawing
Attach the shop drawing PDF and ask: "What are the footing dimensions on sheet S-2? Measure them from the linework and check them against the written dimensions."
Let it draw the sheet
"Make me a plan view of the site at 1:200 with the property lines dimensioned." Open a sheet first, or ask from the scene and it creates one and opens it for you.
Have it show you
"Show me the boundary near the creek" selects it and flies there; "take me back to where I was" steps the view history.
Ask about a library before installing
"Which irrigation libraries are rated highest, and what do the reviews say about the top one?" It can also post, change or delete your own review of a library when you ask.
Iterate on solids
"Make the walls thicker" or "Add a window opening on the north face" — it sees your solid code and parameters.
Good to Know
- One conversation at a time — The assistant processes one message at a time. Wait for a response before sending the next message.
- Context has limits — Very long conversations are automatically condensed to stay within the context window: the oldest messages stop being sent to the assistant, which usually gets a short summary of them in their place and looks one up when you refer to it. After a pause of more than an hour, a question that does not need the earlier conversation may be answered from a short summary of it instead. Either way, a line above a message marks the place and opens to show the summary. If the assistant seems to lose track, start a new conversation.
- It can’t see the 3D view or the sheet — The assistant works through data, not visuals. Describe what you see, paste a screenshot into the message box, or reference entities by name or ID if you need help with something visual. It can select things and move the camera for you, but it does not see the result.
- Rate limiting — Requests are limited to 20 per minute by default to prevent accidental runaway loops.
- It may offer to note a decision — In the scene context, when a conversation settles a design decision on figures you worked out together (a pad size, a setback, a route), the assistant may offer once to save it as a note annotation placed at the entity it concerns, recording the decision, the key numbers and why. It only creates the note if you say yes, and the note follows your approval setting and can be undone like any other change.
- It answers in your units — Measurements and quantities come back in the unit system set by the Metric/Imperial toggle in the status bar, with no need to ask. A figure quoted from a source that uses its own units, such as a survey record, a drawing, a pipe size or a code requirement, keeps those units, sometimes with the converted value in parentheses.
Continue Learning
| Solids & Components | Parametric 3D objects the assistant can help build |
|---|---|
| Terrain | Elevation data, contours, and grading |
| Import & Export | The assistant can handle file operations too |
| Core Concepts | Sites, scenarios, entities, and coordinates |