Using the Athos AI Assistant

Athos is the AI assistant built into VRGS. It is a chat panel you can ask questions in plain English: how to use a feature, what a geological concept means, what is in your project database, or what is currently on screen. Athos understands the context of your project and can search your documents, run database queries, and control the 3D view on your behalf.

By default Athos is cloud-based — it sends your questions to a hosted language model, so it needs an internet connection. It can instead run against a local Ollama server on your own hardware, which keeps your prompts on your machine and works offline. See AI & Machine Learning Requirements for connectivity and account details.

Opening Athos

Athos lives in a dockable panel. To show or hide it:

• On the Home ribbon tab, in the Windows group, click Athos.

The panel opens on the side of the VRGS window. Drag its title bar to dock it elsewhere or float it, like any other VRGS panel. Type your question in the box at the bottom and press Enter to send it. Athos streams its reply into the conversation as it is generated.

When a question needs work first — a database query, a look at the view, a segmentation — Athos narrates what it is doing as it goes, so a long answer shows its progress rather than sitting on a spinner. The final answer then streams in as it is written.

What you can ask Athos

Athos automatically decides which of its abilities to use based on your question. You do not need to tell it which tool to run.

How-to and help questions

Ask how to do something in VRGS and Athos answers from the built-in user documentation (the same content as this site). This help library ships with VRGS, so these questions work out of the box.

"How do I import an LAS well log?" "Where is the flight path timeline?" "What does the M3C2 comparison tool do?"

Geoscience and literature questions

Athos can answer scientific and geological questions from a literature library of your own papers and reports. This searches a document database attached to your project (see Setting up the literature library below). When no library is attached, Athos still answers from the language model's general knowledge, but it cannot cite your specific papers.

"What controls the geometry of hybrid event beds?" "Summarise the methods used to measure palaeocurrent directions in outcrop."

Questions about your project data

Athos can query your project database and answer in plain language, turning your question into a database query for you. Ask about the objects, measurements, and geometry in the open project.

"How many interpretation objects contain the word sandstone?" "List the dip and azimuth measurements grouped by object."

Driving and describing the 3D view

Athos can move the camera, describe what is currently displayed, and take a screenshot of the 3D view and look at it — so you can ask about what is on screen rather than describing it.

"What objects are visible right now?" "Centre the view on the model." "What am I looking at?" "Is the highlighted surface in front of or behind the fault?"

You never ask for the screenshot yourself. Athos decides a question needs a look at the scene and captures it as part of answering, so the picture always reflects the view at the moment it replies — including any camera move it made first.

Along with the image, Athos receives the capture size, how many objects are displayed, how many are selected or highlighted, and the names of the selected ones. That is why it can answer "what is the highlighted one called?" precisely: the name comes from your project, not from reading the picture.

Seeing needs a model that can read images

The hosted OpenAI and Anthropic Claude providers both receive the screenshot. A local Ollama model receives it too, provided the model you are running is vision-capable — the default qwen3-vl:8b is. Run a text-only local model and Athos answers from the scene description alone: it knows how many objects are shown and what they are called, and nothing about how they look. See Images and vision models.

A few things worth knowing about the screenshot:

• It is a rendering, not a photograph. Athos sees your scene with its current lighting, colour maps, shading, and any draped imagery. Colour in a 3D view usually comes from an attribute map rather than from rock, so treat geological readings of a viewport screenshot with caution — for real outcrop appearance, work from photographs.
• The capture is bounded by your 3D window. A larger or maximised 3D view gives Athos more detail. If a feature is off-screen, occluded, or tiny, Athos sees it that way too — navigate to a clear view before asking.
• Each question captures afresh. A follow-up such as "and what about the left-hand side of that image?" will not refer back to the previous picture, so repeat the relevant part of the question.
• A 3D view must be active. If the active window is a photograph, map, chart, or log, there is no 3D scene to capture and Athos will say so. Click into a 3D window and ask again.

Segmenting a photograph

Athos can run the Segment Anything (SAM) tools on a photograph for you. Describe the feature you want outlined and it places the prompts — a box around it, points on it, points on what to leave out — and runs the segmentation.

"Outline the sandstone bed in the middle of this photo." "Segment the scree at the base of the cliff, but not the vegetation." "That mask has taken in too much shadow — exclude the dark area on the left."

The result appears as the ordinary mask preview in the photograph view, exactly as if you had placed the points yourself. Athos never writes to your project: it proposes the mask, and you commit it — or discard it — with Label from Segments, the same as any hand-drawn segmentation. Every mask that reaches your segmentation layer has passed through you.

Athos replies with a summary of what it produced — the area in pixels, the bounding box, and the fraction of the image covered — so you can judge a mask from the numbers as well as by looking at it.

To refine, just say what is wrong. Athos calls the tool again, feeding the previous mask back as a starting point and adding your corrections, so the outline converges rather than starting over.

The photograph must already be open

Athos segments the image currently shown in a photograph view. It will not open a photograph or step to another one — switching images discards the encoded image tiles, which is slow and would take the view out from under you. Open the photograph you want first (double-click it in the project tree, or use Open Photos in Annotation View on an SfM workflow), then ask.

A few practicalities:

• The first request on a fresh photograph warms it up. SAM encodes the image in tiles before it can segment, which takes a few seconds per tile. Athos replies that it is preparing and asks you to try again — the same as the first click of a manual segmentation. Ask again once and it runs.
• The SAM models must already be installed. If they are not, Athos says so rather than starting a download mid-conversation. Click once in a photograph with a segmentation tool to get the usual download prompt, then return to Athos.
• Very small images are refused. Photographs under 1024 px in either dimension are declined rather than segmented inaccurately.
• Refinement needs a hosted provider. Single-shot segmentation works on any provider that supports tool-calling; the back-and-forth refinement loop needs OpenAI or Claude.

See Texture Attributes for carrying committed labels through to the mesh.

Working with photogrammetry (SfM) models

If you build 3D models from photographs with Structure-from-Motion, Athos can help you understand and tune a reconstruction. The quickest way in is to right-click an SfM workflow in the project tree and choose Ask Athos about this Model — this opens the Athos panel and asks it to diagnose that specific reconstruction. You can also just ask in the chat.

"Why does my dense cloud have holes?" "How good is this reconstruction, and which photos failed to register?" "What settings did this use — should I change any?"

Athos reads the reconstruction's settings, its registration and reprojection-error figures, how it is georeferenced, its ground-control and scale-bar accuracy, the worst photos, and the per-view dense-MVS diagnostics — so its answers cite your actual numbers rather than guessing. It also knows the memory cost of the heavier settings, so it steers you away from changes that would run feature extraction out of memory.

Athos can also act on a reconstruction, always with a confirmation step first:

• Change a setting — it shows the exact before → after (for example "depth tolerance 0.025 → 0.040") and only applies it once you agree. The change takes effect on the next reconstruction.
• Run a reconstruction — it will (re)run a workflow when you explicitly ask and confirm. Reconstruction is a long job, so it runs in the background: watch the progress bar, and ask Athos to check the results when it finishes.

Acting on a model needs a hosted provider

Asking about a model and the Ask Athos about this Model menu work with any provider. Changing settings and starting a reconstruction use the tool-calling that the hosted OpenAI and Claude providers support; a local Ollama model can inspect and advise but not act. Switch with /switch claude if needed.

Writing up an analysis

Athos can build an analysis report for you — "write up the stereonet analysis for this outcrop" is a request it can act on. It can create a report, append a heading and body text as a draft for you to edit, and insert an analysis figure captured live from the stereonet, the rose diagram, the Charts panel, a DFN validation run or the 3D view.

Inside a report it also helps with the text itself: checking spelling and grammar, improving, shortening or expanding a selection, describing the figure the caret is in from the statistics behind it rather than from the picture, and reviewing the whole report. Every suggestion is shown and applied only if you accept it — nothing is written into the document on its own.

Sources and references

When Athos answers from your literature library or the help documentation, it lists the material it drew on under a References heading at the end of the reply. Literature sources are shown with their authors, year, title, and DOI where available; help pages are shown with their title, section, and a link. This lets you trace an answer back to its origin and read further.

Choosing the language model

By default Athos uses a hosted OpenAI model. You can switch the provider at any time from the chat box:

Command	Provider	Can see images?
/switch openai	Hosted OpenAI model (default)	Yes
/switch claude	Anthropic Claude	Yes
/switch ollama	A local Ollama server on your own hardware — see setup	Yes, with a vision model (the default qwen3-vl:8b is one)

Use /provider to list the available providers. Running models locally with Ollama keeps your prompts on your own hardware and works offline, but the quality and speed then depend on the model and machine you run; see Running Athos Locally with Ollama to set it up, and AI & Machine Learning Requirements for how it differs from the GPU-based AI tools.

Every provider can answer help, literature, and project-database questions, and all three can now see the 3D view — for Ollama, as long as the local model reads images. What still separates them is the acting: changing settings and segmenting photographs need the tool-calling the hosted providers support.

Slash commands

Type / in the chat box to see command suggestions. The most useful commands are:

Command	Alias	Action
/help	/h	Show the full command reference
/clear	/c	Clear the chat history
/reset		Reset the conversation context (start a fresh topic)
/stop	/s	Cancel the request Athos is currently working on
/export	/e	Export the chat transcript to a file
/switch		Switch the language model provider, e.g. /switch claude
/provider		List the available providers
/schema		Describe the project database schema
/tables		List the tables in the project database
/sample <table>		Show a few rows from a table, to see what is in it
/limit <n>		Cap the rows a database answer returns (/limit off for no cap, /limit to show the current setting)
/sql on | /sql off		Show or hide the database queries Athos runs (/sql always includes queries that returned nothing, /sql status shows the current setting)
/status		Show a system status summary
/version		Show the Athos version
/snapshot	/snap	Save the conversation and its context to a JSON file, with secrets redacted — what to attach to a support report

/help lists everything, including the diagnostic commands (/test_network, /api_count and the various /test_* flows) that are there for troubleshooting rather than everyday use.

Setting up the literature library

The scientific-literature search (the ability to answer geoscience questions from your papers) works over a small search database stored inside your project, in a folder named RAG_DB:

<project> Data\RAG_DB\rag.db
<project> Data\RAG_DB\rag.faiss

These two files are built ahead of time from a folder of PDF papers and reports using the separate RAG DB Creator tool, then copied into the project's RAG_DB folder. Once they are present, restart VRGS (or reopen the project) and Athos will search them when you ask a geoscience question, citing the matching papers in its References.

note

If the RAG_DB folder is empty, geoscience questions are still answered from the language model's general knowledge, and how-to questions keep working because the help documentation is bundled with VRGS. Only the ability to cite your literature depends on the library being present.

Getting good answers

• Ask one focused question at a time. Follow-up questions keep the context of the conversation, so you can refine an answer rather than re-asking from scratch.
• For data questions, name objects and attributes the way they appear in your project so Athos can match them.
• For questions about the 3D view, frame the shot first — Athos sees exactly what is on screen.
• Use /reset when you switch to an unrelated topic — it clears the running context so earlier questions do not colour the new answer.
• If a reply is heading the wrong way, press /stop and rephrase.

Related

• Asking Athos Well — how to phrase a question, what Athos can and cannot see, and what to do when an answer is wrong.

• AI & Machine Learning Requirements — connectivity, accounts, and how Athos differs from the GPU-based AI tools.

• Display Options — control what is shown in the 3D view that Athos can describe.

• Analysis Reports — the documents Athos can draft, illustrate and review.