ouRGeo
Open-Source Spatial IDE
Home Features Docs Download Forum News Plugins Contribute Privacy Sign in Account Sign out

AI Assistant

Providers, thinking, the context the assistant sees, what it remembers, and the controls that keep it to what you allow.

01 The assistant panel

A chat panel that knows your R session — and, when you let it, can drive it.

Open it with the AI Assistant button in the toolbar. Enter sends and Shift+Enter starts a new line. Everything the assistant runs is echoed to your R console like any other action, so it is inspectable, editable and recordable.

Conversations
Keep as many as you like and switch between them from the header. Chats name themselves from your first question and can be renamed or deleted. A saved project keeps its chats and carries them inside the .ourgeoproj; in scratch they last until you leave it.
Model selector
Switch provider and model without leaving the panel. Each chat remembers the model it was held with and restores it, so a conversation written by a large cloud model is never quietly continued by a small local one.
Thinking
Appears for models with supported reasoning controls. See Thinking controls.
Context
How deep the assistant searches your workspace for evidence: Efficient, Balanced or Thorough. It changes retrieval depth, never permissions. See What the assistant sees.
Autopilot
Off by default. Off, the assistant writes code for you to review; on, it creates, modifies and runs R objects in your live session directly. A warning explains the trade-off before it is enabled.
Web
On by default. Off, the assistant has no web tool at all, and its only remaining route to the internet is R code you watch execute. It does not change where the conversation itself is sent — that is your chosen provider.
Terminal
Ships off and works only with Autopilot on. Every command the assistant runs is typed into a terminal tab you can watch and written to the R console.
Attach context / Add files
Stage a layer schema, a console run, a document section or a project file for the next message. See Attaching context.
Dictation
The microphone beside Send. Transcribed on your own machine by default; cloud services are optional, and each upload is logged to the console.

What it can do

Beyond answering questions it can operate ouRGeo itself — open panels and dialogs, add layers, run toolbox tools (through the same data checks you get), write and run R Markdown, arrange and dock panes, build themes, and manage Project Files. It can walk you through a workflow in your own session, and point at a control and wait for you to click it rather than pressing it for you.

Actions are tiered. Navigation and reading are always allowed; anything that touches the R session or the filesystem needs Autopilot; destructive ones also ask you to confirm. Move any action between tiers from the gear beside the Autopilot switch.

The AI panel with Autopilot enabled, streaming R code inline.
The assistant with Autopilot enabled — it writes and runs R, and every line it runs reaches the console.

02 Providers and models

ouRGeo is not tied to one vendor. Choose in Settings → AI Assistant → Provider & model, which also holds the model’s context window — lower it for a small local model, raise it for a large hosted one.

Cloud APIs
OpenAI, Anthropic and Google Gemini with your own API key. AI calls are proxied through ouRGeo’s main process, never made directly from the interface.
Local models
Anything that speaks a compatible API on your own machine — Ollama, LM Studio, llama.cpp, Jan, LocalAI, vLLM, SGLang. Nothing leaves your computer. The health check can set up a local server with a Qwen model sized for your memory and storage. Connections to servers on this computer stay on the loopback interface, with no reply deadline.
Claude subscription
Use a Claude Pro or Max plan through the Claude Code command-line tool you have already installed and signed in to. There is no key to paste and ouRGeo never sees a credential; usage counts against your plan.
OpenAI subscription (ChatGPT)
Continue with ChatGPT opens browser authorisation for an eligible account or workspace; choose a model from its eligible catalogue. Manage usage and credits opens ChatGPT’s usage settings. Credentials stay in the main process and are saved only when a secure system keyring is available. API billing remains a separate provider you choose explicitly, and OpenAI’s eligibility and preview limits apply.
Selected context — your messages, the excerpts the assistant retrieves, the memories it applies — goes to the provider you chose. With a local model, that is your own machine.

03 Thinking controls

The panel footer shows Thinking when the selected model has qualified reasoning controls. Choose a level before sending, or leave Provider default to let the service decide. The choice is remembered per provider and model on this computer; an API connection and a subscription stay separate even when the model names match.

ProviderWhat the selector offers
OpenAI API and ChatGPT planThe levels each recognised reasoning model supports — o3 offers Low, Medium and High; newer models may add Off, Minimal, Extra high or Maximum. Non-reasoning models such as GPT-4o have no control.
Anthropic APIThe documented effort levels of effort-capable Claude models, with adaptive thinking where supported.
Claude subscriptionThe Claude Code CLI’s effort flag for supported models and the Opus/Sonnet aliases; your plan and CLI govern availability.
GeminiNamed levels on level-capable models. Gemini 2.5 shows a Thinking budget of 1,024, 4,096 or 8,192 tokens; Flash and Flash-Lite can also turn thinking off, Pro cannot.
Ollama (native)Exactly the values your server reports for the model, including On/Off. Hidden if the server reports no capability.
Other local endpointsHidden until their reasoning settings can be reliably identified — an OpenAI-compatible interface alone is not evidence.

Higher effort can take longer and use more tokens; the level is a request, and the provider may adapt it. Changes apply to new messages and their tool follow-ups. Private reasoning is never displayed, and a reply truncated mid tool call shows an error rather than running incomplete arguments.

04 What the assistant sees

It retrieves the context a request needs, rather than sending everything every time — and shows you what it used.

The assistant searches skills, action definitions, current layer schemas, console runs, open documents and project files when a request calls for them. An unrelated request — “generate 100 rows of dummy data” — does not wait on R inspection or carry workspace evidence it does not need. Context, in the panel footer and under Settings → AI Assistant → Learning & context, sets how deep the search goes.

The context summary under each reply

Expand the summary below an answer to see the sources selected for it, their revisions, where the request was sent, its delivery state and the estimated input size. Estimated prompt size is kept separate from the input, output, cache and reasoning usage the provider reports — and a failed or cancelled turn keeps the receipts that were available.

Browse current context sources searches the descriptions of everything the assistant could draw on, available and excluded, without reading their contents. From there you can Exclude source, or allow one again.

Attaching context

Attach context stages a layer schema, a console execution or the current document section for the next message; selected line ranges work in large documents too (up to 1 MB per span). Add files and Choose project files attach project files — a whole file, a PDF page, a section, a sheet range or a line range, up to three per message. Removing an attachment chip leaves the file in the library. Answers that use a document cite it, and clicking the citation opens that place.

05 Memory

The assistant remembers what you tell it to, in the scope you choose, and nothing is applied that you have not seen.

The Memory Library is in Settings → AI Assistant and below the assistant conversation. It keeps three scopes apart:

About you
Personal preferences available across projects and chats — “I prefer metric units and concise answers.” Kept in protected local storage on this computer.
Project
Decisions and instructions for one project — “Use miles for this project.” Saved with the project when you save it.
Conversation
Constraints for one chat. They stay with that chat; saved projects keep them.

Adding a memory

Use Add a memory in any scope and choose Remember about you, Remember on project save or Remember for this conversation. Or just say it in the chat:

Remember that I prefer metric units.
Use miles for this project.

Both are handled by ouRGeo itself and do not run a model. Adding, correcting, moving and forgetting all work offline. “Forget…” opens the library so you choose exactly which memory goes — a model never picks it.

Managing memories

Edit
Correct a memory and choose Save correction; the original stops applying.
Move to…
Move to project, Move to About you or Move to conversation.
Always apply
Direct additions and corrections are always applied by default, and the library shows each pin’s estimated recurring cost. Turn it off to let the memory be retrieved only when relevant.
Forget
Stops the memory applying. The original chat messages are kept.
Review transcript copies to forget
Lists the chat messages a memory came from; remove the ones you select. Save the project to commit the removals.
Copy readable export
Copies memories to the clipboard as readable JSONL, so you can see exactly what is stored.

Learning

Settings → AI Assistant → Learning & context sets how memories are learned, separately for personal and project memory: Remember only when I ask, suggest (the default), or automatically remember narrow, declared preferences and decisions. Suggestions do nothing until you accept them in the library, and expire after 30 days. Learning uses local rules — no paid extractor runs. Learning from scratch conversations is off by default.

Forgetting is local. It removes a memory from future prompts. It cannot recall content already sent to a provider, and old exports or backups may still contain it. Deleting a chat also forgets memories derived only from it.

06 Privacy and sources

Settings → AI Assistant → Privacy & sources decides what the assistant may draw on.

Source exclusions
Every source you excluded from the context browser, with Allow again. Exclusions for saved projects stay on this computer; scratch exclusions end with the workspace. While exclusions apply, earlier assistant answers are left out of future prompts.
Workspace column privacy
Choose Refresh workspace objects, pick an object, and tick the columns whose names and types must not reach the assistant. Column metadata loads on selection, so you do not need to open the Data Viewer first.
Reference folders
Folders you explicitly authorise can supply bounded text sources. Refresh after their contents change.

While an exclusion applies, any assistant action that could bypass it — running R that could read the excluded data, for example — is paused. You can still use the interface and the R console yourself. Already transmitted content cannot be recalled from a provider.

In Project Files, clearing Available to the assistant excludes a file’s name, description and contents the same way.

07 Budgets and usage

Settings → AI Assistant → Budgets sets optional limits, separately from the model’s context window.

Chat budget
Estimated daily and monthly limits across all projects on this computer. Each request reserves its input and output before it is sent, using model rates you review (they expire after 30 days), and reconciles with what the provider reports. They are estimates, not invoice caps; while a chat budget is on, a subscription whose processing cannot be priced is refused.
Maintenance budget
A separate limit for background memory maintenance. Paid background processing starts off, and saving a limit does not switch a provider on.

A request stopped by a limit shows BUDGET_LIMIT followed by which limit was reached. Usage totals show the token receipts available for a conversation or project; protected receipts survive chat deletion and restart, and anything missing is labelled unknown rather than guessed.

08 File analysis models

Settings → AI Assistant → File analysis models gives project documents and images their own provider and model. Use assistant model keeps the usual behaviour; choose another and Apply to use a specialist without changing the chat’s model.

  • A document model analyses the excerpts retrieved for your question; the chat receives both the original excerpt and the labelled interpretation, with its citation.
  • An image model interprets a selected image or PDF page, so a text-only chat model can still work with figures. Image interpretation currently needs a native Ollama vision model.
  • Available models is shared with chat — saved and custom choices, subscription models and models discovered on local servers. They reuse your existing connections and sign-ins; there is nothing else to register.
  • Generate AI descriptions on import and rescan (off by default) writes catalogue descriptions for new or changed files, one at a time, using the selected models. They are labelled with their model and date, never overwrite your own description, and may incur usage. Cancel AI descriptions stops them.

A failed task model reports an error; it never silently falls back to the chat model. Its calls count towards your usage and budgets like any other.

09 Storage, recovery and advanced

Storage & recovery
Back up local memory to a verified, encrypted backup and restore missing preferences as reviewable candidates; retry clean-up after a storage problem; and set Diagnostic evidence retention to 1, 3 or 7 days. Expired evidence keeps its content-free token receipts. Personal memory uses your operating system keyring.
Advanced
Optional memory connections and experimental local engines. A connection can store protected credentials and test reachability without sending chat or memory content.
Experimental, Linux x86_64 only. The local memory engines under Advanced each require an explicit acknowledgement and review before they run. Existing memories are never imported automatically — you select which to migrate — and the built-in Memory Library needs none of them, no account, no embedding download, Python or Docker.

Plugins as context sources

A plugin can contribute skills and Markdown context sources behind a separate context-source permission, and memory profiles behind a memory-provider permission. Enabling a plugin alone shares nothing. Its sources appear in Browse current context sources, where you can exclude them, and an update that asks for new permissions switches the plugin off until you review it. See Plugins.

10 The AI Assistant settings pages

PageWhat it holds
Memory LibraryAbout you, Project and Conversation memories — add, edit, move, forget, review copies.
Provider & modelThe assistant, its connection, model and context window; ChatGPT and Claude subscription sign-in.
File analysis modelsSeparate document and image models, and AI descriptions.
Learning & contextContext effort, and personal, project and scratch learning.
Privacy & sourcesSource exclusions, workspace column privacy, reference folders.
BudgetsEstimated chat limits and the maintenance budget.
Storage & recoveryBackups, clean-up, diagnostic retention.
AdvancedMemory connections and experimental engines.

Voice input is its own page in the Assistant group — see Settings.