The native add-in owns UI, document extraction, preview, and write-back. Rules, templates, configuration, logs, diagnostics, and model calls stay in the local adapter. Current scope is the platform foundation plus three hosts, targeting Kylin V10 ARM, Python 3.8, and offline install.
3Separate host add-ins; Ribbon never cross-displays
8Write, review, analysis, and summary tasks
127.0.0.1Local adapter, default port 18100
3.8Target Python, offline installer
Features
Model output is previewed before any write-back. Review and analysis tasks are read-only by default. Workflow-platform access uses
/chat-messages; direct-model access uses OpenAI-compatible
/chat/completions. Runtime requests do not fall back to a unified URL or key.
Template-based imitation. Preview, plain text, copy; no write-back.
Document Review
Typos, expression, logic, fluency, professionalism. Selection or limited full document.
Format Review
Against the technical-document format rules. Format-issue cards, caption-association conclusions, image-semantics supplement. No format write-back.
Read-only · visual-off degrade
Writing policy
Four preset packs plus a local organization library. Available to write / imitate / review.
Excel
2 entries
Smart Analysis
Selected or used range. Structured report and briefing paragraph; no cell writes.
Read-only
Formula Assistant
Explicit selection, max 30×20. Generate or explain; copy only.
Copy only
PPT
2 entries
Smart Summary
Current slide, or one .md / .docx (≤10 MB) for a full-deck outline. Preview and copy only.
Preview / copy only
Structure Review
Up to 60 slides. Page-role list first, then rule exemptions. Read-only.
Read-only
Production mock output stays off unless explicitly enabled. Image-semantics supplement is default-on; probe failure or a closed master switch is visual-off degrade — never claim the model saw the picture. Automated gates still yield only
candidate.
Architecture
Each layer has one job. Documents travel as structured payloads. Health state decides whether config changes and new model jobs are allowed.
01User in WPS
Works from Word / Excel / PPT without leaving the document.
02JS/HTML add-in
UI, extraction, preview, write-back. Separate packages so Ribbon buttons never cross-display.
03Document bridge
Paragraphs, headings, fonts, sizes, alignment, outline levels stay on the payload.
04Local adapter
127.0.0.1:18100. Rules, templates, config, logs, model calls.
05Enterprise AI API
Workflow platform or direct model. Keys live in local reference files, not in config text.
Preview, then write
Word writing tasks preview first. Review, analysis, and summary do not change the source by default.
Four health states
Live, ready, degraded, recovery. Recovery blocks config changes and new model jobs.
Per-task config
Eight tasks keep their own access method and key references. No unified URL fallback at runtime.
Release notes
Current v0.25.3-alpha. Version rule
AI-WPS-P{phase}-{scope}-{major.minor.patch}-{yyyymmdd}. Major is a compatibility boundary, minor is user-visible capability, patch covers fixes, UI, packaging, and docs.
v0.25.3-alphacandidate · d1a346b
Result preview: restricted Markdown for headings, lists, tables, and bold
Format issues become cards with a separate action bar; caption association is linked / orphan / missing / ambiguous
Current kit
ai-wps-phase1-delivery-20260826-d1a346b-v0253.tar.gz
· SHA-256
120a2cfd…258cb7.
Rejected predecessors, gate numbers, and lineage live in
v0253,
v0252,
v0251.
Install
For local development, start the adapter then load the add-in. Intranet terminals use the offline kit. Overwrite installs keep existing config, keys, the writing-policy database, and backups.
01
Start the local adapter
/health/live does not read business data. /health/ready returns 503 in recovery. Without FastAPI deps: python adapter_service/standalone_adapter.py 18100.
Default output dist-offline/wps-ai-assistant-offline.tar.gz. When system Python has no pip, get-pip runs with -sS so Kylin apt dist-packages are not scanned.
Tests: cd adapter_service && pytest; cd wps-addon && npm test. Target-machine regression uses Python 3.8 on Kylin V10 ARM64. Later work on the same adapter can add multi-sheet Excel flows, multi-file compare, governed PPT generation, and richer policy governance — outside current Phase 1.