Skills

Understand a document capability

Choose a task first. Use an account with access to that feature in the active organisation; each linked procedure lists its prerequisites.

1. Understand shared methods
2. Check availability first
3. Read extraction coverage

Check the result: You can distinguish supported content, source provenance and coverage limits before using an AI result.

If you get stuck: If a format or capability is unavailable, consult the availability section and use a supported source. A published manual does not establish production availability.

What Skills are

Use case: Understand why the same document produces consistent evidence in different Atlas surfaces.

A Skill is a named, versioned method that Atlas resolves by exact identity. Knowledge (My Docs and Company Docs, the two scopes of one Library) and Assistant can use the same PDF understanding contract instead of maintaining separate readers.

A Skill is not a permission grant. Your signed-in role and the service that owns the data still decide what you may read or change. Skill definitions do not contain executable uploads, credentials or connectors.

Expected result: You can distinguish the method used from the authority to use it.

Use PDF understanding in Knowledge

Screen: /knowledge/library/:fileId

Use case: Turn a text PDF, scan or image-heavy PDF into cited knowledge.

Before you begin: Use an organisation role with permission to contribute Knowledge files.

How to do it

1. Open Knowledge → My Docs or Company Docs — both are scopes of one Library — and upload a PDF, then confirm the draft after checking its name and location.
2. Open the ready document. In PDF understanding, check the skill identity, file metadata, inspected pages and separate text, visual and AI coverage.
3. Read passages with their page and source kind. Text came from the PDF text layer, scan is recognised text from a rendered page, and visual describes a figure, chart, table or image.
4. Use Re-extract when you need the current skill version to read the original file again. The previous result remains visible until the replacement is ready.
5. If Atlas reports Partial coverage, read every named gap before relying on the result.

Expected result: The document shows page-anchored passages, metadata and an honest coverage record without editing the original file.

Read a PDF in Assistant

Screen: /assistant

Use case: Ask a question about a PDF attached to the current conversation.

How to do it

1. Start a conversation from AI Assistant → New chat, or the round Assistant launcher, and attach the intended PDF. Add the relevant customer context only when the question needs it.
2. Ask a specific question, such as which warranty term or model appears in the file.
3. Check that the answer cites the relevant page, for example [p.1], and states when information was recognised from a scan.
4. Open or download the original when a decision depends on wording or visual detail. A cited extraction helps you locate evidence; it does not replace the source document.

Expected result: Assistant uses the same versioned PDF method and answers with page-level evidence.

Read provenance and coverage

• Skill identity: the skill ID and version used for this run. The first released capability is document_pdf_extract v1.0.0.
• Metadata: available title, author, creation and modification dates, page count and detected language.
• Inspected pages: pages the run attempted within its bound. Pages outside the bound appear as a gap rather than disappearing silently.
• Text, visual and AI coverage: separate lists show where Atlas found a text layer, rendered visual content and model-assisted understanding.
• Passage provenance: each passage identifies its page and kind; model-assisted passages also identify their model route.

Complete coverage means the bounded inspection finished without a known gap. It is not a guarantee that every statement or visual nuance was interpreted correctly.

Inspect Skills as a platform superadmin

Screen: /platform?view=skills

Use case: Confirm which governed definition and version produced a run.

Before you begin: This page requires the separate platform-superadmin role. Organisation administrators do not receive platform authority from a Skill.

How to do it

1. Open Platform → Quality → Skills.
2. Select a definition to inspect its owner, input and output contract, constraints, model route and active version.
3. Review recent runs by surface and status. Match the skill ID, version and digest to the document or Assistant result you are investigating.
4. Use rated usage to reconcile model tokens and the provider-rate identity. Missing or unrated usage is a gap to investigate, not zero cost.
5. Do not enable, disable or publish a version unless that platform operation is separately authorised.

Expected result: You can trace a user-visible result to an exact governed method without confusing inspection with permission to change it.

Use image understanding

In Knowledge → Library, upload and confirm a JPEG, PNG, WebP, GIF, BMP or safe SVG. The detail shows the exact document_image_extract version, dimensions, literal capture metadata when present, separate transcription and description passages, and coverage. GIF reads its first frame; SVG is checked and rasterized by Atlas, never executed in your browser.

Assistant read_artifact uses the same authorized evidence path for an attached image. Keep the original open when a decision depends on details: image output is untrusted evidence, never person identification or an instruction to follow.

Prepare a video for AI use

For a video larger than 50 MB (50 MiB), select the original video file from your device when prompted, then review the processing confirmation. Atlas checks that it matches the uploaded original. If it does not match, select the original again; uploading alone does not authorize processing.

Before you begin: prepare a video of at most five minutes and use a browser with WebGPU support.

Upload an MP4, MOV or WebM up to 1 GB (1 GiB). It stays in the Library as unprocessed; uploading does not transcribe or profile it.

1. Open the video and choose AI Readiness. Processing supports videos up to five minutes. For longer material, upload a shorter clip.
2. Read the keep-page-open explanation, then start. Your browser downloads the local models on first use. Transcription and frame descriptions run on your device; a capable browser with WebGPU is required for visual profiling.
3. Keep the page open while media preparation, transcription, visual profiling and saving finish. You can cancel. When leaving, choose Stay to keep processing or Leave to stop it; refresh and close warnings are controlled by the browser.
4. Only a fully saved result becomes AI-ready. Reopen it to read the transcript, sampled frame descriptions, timestamps and coverage. Select a timestamp to seek the original video, or search for its recorded words in the Library.

Speech stays in its original language. Frame descriptions are in the model’s output language. Both are browser-reported AI evidence and may contain mistakes; compare them with the original. Sampled frames do not describe every moment or motion. A silent video can have a visual-only result.

If decoding, device capacity or inference fails, the original remains available. Follow the stated remedy or retry on a capable device. Closing the browser does not finish a job in the background. This path uses no paid inference fallback.

Result: after the complete run is saved, timestamped video evidence is available to revisit and search.

Frames are resized to limit device memory. Samples that exceed the size limits are skipped and reflected in coverage.

Availability and limits

Current availability: image understanding is staged with the shared Skills framework. Production availability requires a separately verified product release.

• JPEG, PNG, WebP, GIF, BMP and safe SVG are limited to 8 MiB and 4 million pixels.
• One cache-disabled governed model request produces bounded text and description. A limit or unavailable model is stated as Partial coverage.
• GIF reads only the first frame. Unsafe SVG features and external references are refused before model use.
• Video processing starts explicitly in the Library and requires a capable open browser; results carry browser-reported provenance.