Epic Overview

ACM-AI was implemented across 17 epics covering 122 user stories. As of 2026-02-22, all feature epics are complete (feature-complete status).

Epic Summary Table

Total: 112 stories completed (92%) — all feature epics complete

Epic Descriptions

E1 — ACM Data Extraction Pipeline (31 stories)

The largest and most complex epic. Covers the complete 7-stage extraction pipeline from document structure analysis through to SurrealDB storage. Key milestones:

• E1-S1 through E1-S7: Core data model, domain objects, API endpoints, and source processing integration
• E1-S8: Site configuration data entry for non-extractable fields
• E1-S9: ACM product taxonomy classification (pattern-based + LLM fallback)
• E1-S10: MinerU table extraction with merged cell handling
• E1-S11: Generic configurable parser driven by BAR field schema (replaced 3 consultant parsers)
• E1-S12: Consultant wording normalisation to BAR enum vocabulary
• E1-S13 through E1-S15: Corrective RAG validation, contextual embedding enrichment
• E1-S16 through E1-S19: Document structure analysis, building inventory, page-level section tagging, metadata extraction
• E1-S20 through E1-S21: Agentic orchestrator, pipeline visualisation UI
• E1-S22 through E1-S31: BAR schema expansion, Victorian Government hierarchy, removal tracking

E2 — AG Grid Spreadsheet Integration (12 stories)

Full AG Grid implementation with 47+ BAR columns organised into 7 column groups (Organisation, Building, Location, ACM Details, Assessment, Documentation, Removal). Includes:

• Virtual scrolling for 1,000+ rows
• Column visibility presets (Essential, Full BAR, Assessment Focus, Removal Tracking)
• Row grouping by Building and Room
• Risk status colour coding (Low=green, Medium=yellow, High=red)
• Dynamic column definitions from the field_schema API (E2-S8)
• AG-UI incremental record streaming during extraction (E2-S12)

E3 — Cell Citations and PDF Viewer (4 stories)

Click any cell in the AG Grid to open the source PDF at the exact page. The citation format [acm:record_id:field_name] extends the existing source reference system. Bounding box highlighting (FR-403) is implemented when MinerU provides coordinate data.

E4 — Chat with ACM Context (4 stories)

Supervisor agent pattern using LangGraph and CopilotKit. The ACM context toggle dynamically enables or disables ACM-specific tools (record queries, building search, compliance checks) per message. Chat responses include [acm:...] citations that are clickable in the frontend.

E5 — Export Functionality (4 stories, promoted to P0)

BAR-compliant Excel export (.xlsx) with all 47 columns in the official BAR column order (A–AU). Also includes CSV export, column order configuration, and building-level filter for selective export.

E6 — Rebranding to ACM-AI (4 stories)

VAEA government branding applied across the application: teal/coral/navy colour palette in OKLCH colour space, updated logo and favicon, navigation renamed to WORKSPACE/CONFIGURE taxonomy.

E7 — Upload Wizard (7 stories)

Multi-step upload workflow: file selection → site configuration (Department, Agency, Building Type) → extraction trigger → progress monitoring → register view. Supports batch uploads and configuration templates.

E8 — UI Refresh (Bento Grid) — ARCHIVED

Originally planned bento-grid layout redesign. Archived in favour of the VAEA government branding approach implemented in E14. Code preserved but inaccessible from navigation.

E9 — Document Library Management (3 stories)

Unified Documents view at /documents merging the previous Sources and Documents pages. Includes bulk operations (select multiple → delete/re-extract/export), inline extraction log expansion, and document status badges.

E10 — ACM-AI UI Simplification (1 story)

Removed brownfield features (Podcasts, Transformations, Notebooks) from navigation while preserving code. Features remain accessible via direct URL for backward compatibility.

E11 — Search and Retrieval Enhancement (2 stories)

Parent document retrieval using acm_table_section records as parent context for extracted ACM records. Enables provenance-aware RAG where chat answers can cite the raw table HTML alongside the structured record.

E12 — Extraction Settings and Configuration UI (4 stories)

Settings pages under the CONFIGURE section: parser configuration, field schema management (E12-S4 — BAR Field Schema Configuration UI), model selection, and processing options.

E13 — Knowledge Graph Visualisation (3 stories)

Graph view showing relationships between buildings, rooms, and ACM items. Uses a force-directed layout to visualise the Victorian Government hierarchy (Department → Agency → Site → Building → Room → ACM Item).

E14 — UX and Enterprise Readiness (11 stories)

All 11 stories implemented and merged via PR #7 (2026-02-08). Covers:

• VAEA design token system (3-tier cascade: Brand → Semantic → Component)
• Skeleton loading placeholders (zero CLS, shimmer animation, aria-busy)
• Toast notification system (Sonner, promise-based for long operations)
• WCAG 2.1 AA compliance audit and remediation
• Breadcrumb navigation on deep pages
• TypeScript type generation from Pydantic models

E15 — Extraction Monitor and Live Logging UI (2 stories)

• E15-S1: Inline ExtractionProgressPanel in Document Library (chevron expand, historical/live SSE)
• E15-S2: Dedicated /extraction-monitor page with Active and History tabs, status filter, retry capability

E16 — UX Enhancement Sprint (3 stories)

• E16-S1: Statistics dashboard home page (4 summary cards, risk donut chart, top-10 buildings bar chart, recent extractions list)
• E16-S2: Slide-out record detail panel (380 px right drawer, 8 field sections, keyboard navigation, edit mode)
• E16-S3: Empty state screens with onboarding hints on first visit

E17 — Live Extraction Intelligence — AG-UI + A2A + Observability (6 stories)

The final feature epic (2026-02-22):

• E17-S1: AG-UI extraction SSE endpoint using SurrealDB relay (agui_events table, migration 21)
• E17-S2: Incremental record streaming to AG Grid via StateDelta events (preview rows with italic/pulsing border)
• E17-S3: Reasoning token display panel for thinking models (DeepSeek R1, Claude extended thinking)
• E17-S4: Live tool call feed showing step names, arguments, results, and durations
• E17-S5: A2A agent card at /.well-known/agent.json with task lifecycle endpoints
• E17-S6: 6 new frontier AI models via OpenRouter (auto-provisioned when OPENROUTER_API_KEY is set)