ConstruChain Vision

Inspect a model's property data structure before writing SQL. Returns the table's actual columns, total_rows, total_objects (distinct elements) and the available sections with their property names and row counts. The data is a single LONG-format table named `properties` (one row per object+property; values are VARCHAR). If the property list is truncated, drill down with: SELECT DISTINCT name FROM properties WHERE section = '<section>'

get_model_schema

List the BIM models in your workspace that have property data ready to query. Returns one entry per model: model_guid (use it in the other tools), model_name, project_name, file_type, parquet_size_bytes, last_imported_at. Always call this first to discover the model_guid you need.

list_models

Publish a self-contained HTML dashboard so the user can view it in ConstruChain Vision and optionally share it. If you pass model_guid, the live 3D viewer is added automatically beside the dashboard. Drive it ONLY via window.VisionViewer.isolate/color/zoomTo with property filters {section, name, values} — never call parent/top.postMessage yourself and never send raw object IDs/dbIds (publish is REJECTED, and even if it weren't the viewer ignores them). See the "Vision 3D viewer" note below. Build the HTML under these constraints (violations are rejected): - One complete, self-contained document. All data INLINED as a static snapshot (e.g. a const in a <script>). The page runs sandboxed with NO network access — fetch/XHR/WebSocket all fail, so never load data or assets at runtime; loaders that fetch a file (GLTF, d3.json, map tiles) will not work. Inline geometry/data instead. Scripts run only after the complete HTML has streamed, so incomplete/truncated documents are rejected. There is no chunked/partial upload — the whole document must arrive in one call, so keep it compact (see the size note below). - Use any of the hosted libraries below, referenced from /assets/. Combine only the few you actually need (referenced libs do NOT count toward the size limit, but each one is extra weight the viewer loads). No other external scripts, styles, fonts or CDNs — only /assets/ is reachable, and no library here needs unsafe-eval. Classic globals — add <script src="/assets/NAME"></script>, then use the global: charts: echarts-5.6.0.min.js (echarts), chart-4.4.9.umd.min.js (Chart), apexcharts-3.54.1.min.js (ApexCharts), plotly-2.35.2.min.js (Plotly), d3-7.9.0.min.js (d3; also d3.format for number/currency), uplot-1.6.32.iife.min.js (uPlot, dense time-series; + <link> uplot-1.6.32.min.css) 3D: babylon-7.34.4.min.js (BABYLON) UI (React-like, no build): preact-10.24.3.umd.js (preact) + preact-hooks-10.24.3.umd.js (preactHooks) + htm-3.1.1.umd.js (htm) tables: tabulator-6.3.0.min.js (Tabulator; + <link> tabulator-6.3.0.min.css), gridjs-6.2.0.umd.js (gridjs; + <link> gridjs-mermaid-6.2.0.min.css) animation: gsap-3.12.5.min.js (gsap), anime-3.2.2.min.js (anime) dates: dayjs-1.11.13.min.js (dayjs) Vision 3D viewer: pass model_guid to link the model and the trusted dash shell renders the live 3D viewer AUTOMATICALLY in a resizable split pane beside your dashboard (identical to the in-app viewer, with a draggable divider). DO NOT add any viewer element to the HTML — no <vision-model-viewer>, iframe, canvas, Autodesk/Forge client, and no split/column/grid layout reserving space for 3D. Build the report as if it fills the full width; the viewer is added to its right. Drive it from your charts/buttons via window.VisionViewer (the vision-viewer-bridge.js script is injected automatically for model-linked dashboards): VisionViewer.isolate( filter); VisionViewer.color(filter, '#2563EB'); VisionViewer.zoomTo(filter); VisionViewer.reset(); VisionViewer.clearColors(). A filter is {section, name, values:[...]} — property filters only. Example (query_model first to get the exact section/name/values, then wire a button): <button id="op">EM OPERAÇÃO</button> <script>op.onclick=()=>VisionViewer.isolate( {section:'SmartBIM',name:'STATUS',values:['EM OPERAÇÃO']});</script> Never call parent/top.postMessage yourself and never send raw object IDs/dbIds, Autodesk URNs, access tokens, or any network call from dashboard HTML (all are rejected at publish; the viewer accepts only property filters). filter.values must match the model's stored text EXACTLY (case, accents and even source typos included) — query the real values first with query_model/get_model_schema; a value that does not match simply isolates nothing. There is no filter-by-object_id command by design: to target a specific element, inline a map from object_id to a per-element unique property (UID/GlobalId/Tag/…) and isolate by that property. Live filtering needs the model's processed property index; if commands do nothing and the viewer reports "property artifacts are not available", the model must be re-processed once (Rerun Processing). Cross-filter contract: combine active dashboard dimensions and status in one filter array. Separate filter objects are intersected (AND); values inside one filter are alternatives (OR). When an interaction colors and selects, send color commands first with the same active dimension filters, then finish each interaction with isolate or zoomTo. Final isolate/zoomTo result becomes toolbar selection used by "Zoom on selected" and "Hide unselected"; color never defines toolbar selection. Example: const selection=[ {section:'SmartBIM',name:'MODULE',values:['BOX1-MD03']}, {section:'SmartBIM',name:'Status',values:['AVAILABLE','BUSY']} ]; VisionViewer.color([ selection[0], {section:'SmartBIM',name:'Status',values:['BUSY']} ], '#EF4444'); VisionViewer.isolate(selection); ES modules — declare an import map, then import in a module script: <script type="importmap">{"imports":{ "three":"/assets/three-0.169.0.esm.min.js", "lit":"/assets/lit-3.2.1.esm.min.js", "lit/decorators.js":"/assets/lit-3.2.1.esm.min.js"}}</script> <script type="module">import * as THREE from 'three'; // WebGL/3D import {LitElement, html, css} from 'lit'; // web components </script> - Fonts: @font-face with src url(/assets/inter-variable.woff2) — also montserrat-variable.woff2, jetbrains-mono-variable.woff2, geist-variable.woff2. No Google Fonts / remote fonts. - CSS: inline (<style> or style=), or a hosted <link rel="stylesheet" href="/assets/…css"> for the libraries above. Images only as data: URIs. - Do not include ConstruChain/Vision product chrome: no topbar, navbar, report button, legal banner, or shell around the dashboard. The dash host adds the standard Vision topbar above your HTML; start your document at the dashboard content. - Not allowed: iframe/object/embed/form/base/applet, <meta http-equiv>, srcset, and any external or relative URL (only '#...', 'data:' and '/assets/...' pass). Inline <script>/<style> and import maps are fine. - Max size 1.5 MB — aggregate with query_model instead of embedding raw rows. - Charts inside hidden tabs/panels (display:none) often initialize at 0x0. Initialize them lazily when the tab becomes visible and call the chart library's resize() method after showing the panel. visibility may be "workspace" (default), "members", or "public". "members" requires model_guid (from list_models) or project_id so Vision can link the dashboard to a project; Phase 1 treats members as workspace members, and Phase 2 tightens it to project access. Public dashboards are shareable by URL, so use them only when the user explicitly asks. Workspace dashboard limits come from the user's current plan. By default the dashboard is created with workspace visibility: workspace members view it in Vision (manage_url). Only set visibility="public" when the user explicitly asks for a shareable public page. To update an existing dashboard with fresh data pass its dashboard_id (URL and visibility are preserved). Pass dry_run=true to validate the title/html and receive errors without creating or replacing anything. For a large dashboard, aggregate the data with query_model first (never embed raw rows) and run a dry_run to catch truncation/size before spending the publish.

publish_dashboard

Run a read-only DuckDB SQL SELECT against one model's property data. The data is exposed as a single table named `properties` in LONG format (see get_model_schema). Full DuckDB SQL is supported: CTEs, window functions, PIVOT, regexp_matches, TRY_CAST. Values are stored as VARCHAR — use TRY_CAST(value AS DOUBLE) or regexp_extract before numeric aggregation. Units and date formats come from the source model and may vary by discipline, so inspect distinct values before labeling units or dates. Exactly one SELECT/WITH statement is allowed. Returns {columns: [{name, type}], rows: [[...]], row_count, truncated, elapsed_ms}. When truncated is true the row or byte cap was hit — add GROUP BY/filters/LIMIT and retry instead of paginating raw rows.

query_model