Match a P&ID with its extracted knowledge graph as a graphical overlay. Inspect the knowledge graph (BOT / FSO / Brick, Turtle) directly on the drawing it came from, and on every other artifact it was built from — detail sheets, plans, BMS screenshots, catalog pages, product sheets and point lists. Run SPARQL on the graph (also from plain words) with the results shown on the drawing. Refine the graph with issue reports, review rules, edits to instances and relations, or by having Claude re-check sub-areas and resolve issues. Fixes go back into the graph through a change log, and the refined graph exports as Turtle together with a SPARQL Update.
- Online app (runs inside Claude): Graph Overlay Inspector — https://claude.ai/artifact/CdtPc1qnDAVkdRgjQKuU3R, with the HTRK-VEN03 ventilation demo preloaded.
- Single-file copy:
graph-overlay-inspector.html. It works in any browser, also offline (except PDF rendering); Claude features there use your own API key (see §1). - Your copy is your own. A copy from GitHub runs entirely in your browser: what you load, edit and save stays in your browser's storage and your own files, and nothing reaches the online app or anyone else's work. Claude features in the copy are billed to the API key you enter (it stays in your browser and is sent only to api.anthropic.com). Open in Claude links to the author's online app, which only opens for people it has been shared with; inside Claude, Claude features run on the viewer's own Claude account (each viewer is asked first) and each viewer's work stays in their own browser.
- Several artifacts, one graph: a project holds any number of sheets and point lists, in groups (domains). Every sheet is a view of its own with side diagrams around it; the Evidence panel and the Instances table show where each instance appears (see §3).
The ventilation demo: every instance of the graph on the P&ID it came from, its links drawn between them, other artifacts as side diagrams, and the review checks on the left.
Choosing an instance shows its links on the drawing and its card: classes, system, location, every connection, and owl:sameAs names from other artifacts.
SPARQL on the graph, with every hit ringed on the drawing.
Editing an instance by hand: every triple in and out of it, with warnings before a critical link changes.
- It is a page hosted on claude.ai, not part of a chat. It keeps working after the chat that built it ends. Every republish updates the same link for everyone.
- Shared as Anyone with the link (from its Share menu), everyone can open it. Loading files, overlay, review, checks, edits, applying fixes, SPARQL and exports work for everybody.
- The Claude features (Anchor, Re-check areas, Resolve with Claude, Suggest placements, SPARQL from plain words) work when the link is opened inside claude.ai while signed in. The calls run on the viewer's own Claude account and usage, never the author's. The first call asks the viewer to allow it, including sending images. If the viewer's plan lacks Claude Opus 5.5, a nearby model answers; the run log names the model.
- Every person's work is private to their own browser. A review travels as a project file.
- Only the author can change the published app. Anyone can make their own copy (chat or HTML file); the original and its link are not affected.
- Saving files inside Claude: Claude lets a page save only some file types (images, txt, json, md, csv, html, svg, pdf, docx, pptx, xlsx, zip …).
.ttlis not among them, and the page cannot change that. When a.ttlsave is refused the app offers: Try direct .ttl (an ordinary browser download, which the sandbox may block), the refined graph package (.zip), Save as .txt or Copy. Outside Claude the.ttlis saved directly.
Everything except the Claude features works locally, and .ttl exports download directly. For the Claude features:
- Your own Anthropic API key (Open → Claude API key (.env)…, the setup card, or the Open in Claude dialog):
- Enter the key, choose the model (Claude Opus 5.5, Sonnet 5.5, Haiku 5.5 or Fable 5.1) and the max output tokens. Test sends a one-word request.
- The app calls
https://api.anthropic.com/v1/messagesdirectly from the browser (streaming). Usage is billed to the key's account. Images are scaled to about 1.15 MP / 1568 px first. - A
.envfile (named.envorclaude.env):ANTHROPIC_API_KEY=sk-ant-… CLAUDE_MODEL=claude-opus-5-5 CLAUDE_MAX_TOKENS=32000 # only for keys not scoped to one workspace: ANTHROPIC_WORKSPACE_ID=wrkspc_…CLAUDE_API_KEY,ANTHROPIC_MODEL,MODEL,MAX_TOKENSandWORKSPACE_IDare accepted too. Save as .env file writes one for you. - Workspace ID: a key that is not scoped to one workspace needs it, or the API answers 400. It is sent as the
anthropic-workspace-idheader; the dialog has a field for it and says so when the API asks for it. - Where the app finds the file: served over http(s) (e.g.
python3 -m http.serverin that folder) it is read next to the app by itself. Opened from disk (file://), browsers don't let a page read its neighbours — use Load .env file…, or open or drop a folder that contains it, and tick Remember in this browser if you like. - The key stays in the browser (only in
localStorageif you tick Remember; Forget key removes it), is sent only to api.anthropic.com and is never written to project files. Never commit a.envfile;.gitignoreexcludes them.
- Open in Claude: save the project file (
.project.json: graphs, drawings, images with notes, point lists, placements, review, rules, proposals, change log) and drop it on the online app.
Limits of the file route: the hand-off to Claude is manual (save + drop); PDF drawings need internet (pdf.js 4.10.38 loads on demand); the demo is not in the file (serving site/ over http includes it).
- Start a review. Drop the graph and the artifacts — files or whole folders — or use Open.
- Graphs: Turtle / N-Triples / TriG (several files are merged). A new version of a loaded graph (same name, or most of the same entities) is compared with it before it replaces it: you tick which loaded file(s) it replaces (the others stay), see every entity whose triples differ with version 1 (current) and version 2 (new) side by side per predicate, and keep v1, v2 or both — per value, per entity or for all. Placements stay where the IRI is unchanged; your review's changes are re-applied; what you keep from the old version is one change-log entry, and the replaced file is not kept in the project. Artifacts: PNG, JPG, WebP, GIF, BMP, SVG, TIFF (scans, decoded in the browser) and PDF pages; point lists as
.csv,.tsvor.txt; a project description as.md/.txt. - Folders become groups (domains such as Ventilation or Heating): the first folder under the one you open. The open dialog shows a block per group; say for each file what it is — P&ID / diagram, Plan, BMS screenshot, Catalog page, Product sheet, Functional report, Description, Other sheet, or Legend, Title block, Context image — and add a note for Claude. Dropping more files while the dialog is open adds them to its list. A file the browser can't read is listed and left out; the rest opens.
- Project description (optional): the Description tab in the sheet bar, the setup card or Open → Project description…. Claude reads it with every request.
- Graphs: Turtle / N-Triples / TriG (several files are merged). A new version of a loaded graph (same name, or most of the same entities) is compared with it before it replaces it: you tick which loaded file(s) it replaces (the others stay), see every entity whose triples differ with version 1 (current) and version 2 (new) side by side per predicate, and keep v1, v2 or both — per value, per entity or for all. Placements stay where the IRI is unchanged; your review's changes are re-applied; what you keep from the old version is one change-log entry, and the replaced file is not kept in the project. Artifacts: PNG, JPG, WebP, GIF, BMP, SVG, TIFF (scans, decoded in the browser) and PDF pages; point lists as
- Place instances on the sheets. Coordinates in the graph (numeric
…X/…Ypairs) are used automatically. Anchor with Claude places the rest — on this sheet or on every sheet in turn. "Only instances not placed on this drawing" counts the same instances as the scope line: never placed, plus those an earlier run marked not on this drawing (Look again retries them). The result is a table per sheet (placed / not on it / what happened); Recently anchored above the drawing then shows only what that run placed, with no lines until you click one (Esc hides them again, Show everything goes back). Place on another sheet… in an instance card, and the Instances table for the ones on no artifact yet (place by hand or let Claude suggest). - Review. Inspect, drag, decide (Correct / Issue / Unsure / Resolved, with comments), decide whole reports at once in Review & checks, flag links, make one issue for several instances and links, check point-list rows, propose or apply relations, rename instances and edit labels and comments.
- Re-check areas marks doubtful places on a drawing; Claude compares them with the graph in one call.
- Issues → Resolve with Claude sends the open flags, rules, notes and group issues you pick — with editable comments, bundles and an overall comment — in one call. Claude proposes concrete fixes.
- Accept the fixes you agree with, then Apply to graph. Each applied fix becomes a change-log entry.
- Query the graph with SPARQL — or describe what you want and let Claude write the query.
- Export the refined graph (.ttl), the refined graph package (.zip), the SPARQL Update, the change log, the review report, the instance coverage table or the project file.
Undo / redo (↶ ↷, Ctrl/⌘+Z, Ctrl/⌘+Shift+Z) covers moves (also on side diagrams), placements on any sheet, guests, reviews and per-report decisions, link flags, rules, group issues, proposals, renames and text edits, relation changes, time series IDs, point-list edits, Claude runs, accepting, applying and reverting. It holds 200 steps.
- Click an instance: it is chosen and the view zooms to fit it and its connections (switchable in View settings). The selection's lines are drawn on top of the labels; switched-off categories show dimmed and dashed. Clicking the empty drawing lets go and zooms back.
- Ctrl/⌘-click adds or removes instances — without zooming. Ctrl/⌘-click a line adds that link (predicate) to the selection; chosen links get a halo.
- Box select: Shift-drag on empty space, or switch on box select (dashed square by the zoom buttons) and drag; Ctrl/⌘ adds. Optionally the box also picks the links between the instances inside it. Dragging with the mouse wheel button always pans.
- Instance card: class, system, location, connections (every asserted triple, in its direction), Time series for points, Properties (all literal values, links to things that are not drawn, and Referenced by: triples pointing at the instance from things that are not drawn — a product class, a type …; click one for everything the graph says about it), owl:sameAs twins, placement and where it appears, its issues and checks, and its review. Edit… next to the IRI renames it, edits labels and comments, and edits every triple in and out of it (blank nodes included).
- Select similar ▾ (on an instance card): instances with the same class, the same set of classes, the class or its subclasses, the same system or location, the same name or label pattern (digits as wildcards), the same kinds of links or the same properties — each choice shows how many.
- Classes (top bar, crown icon): every class the graph itself describes is an entity — inspect it (superclasses, subclasses, instances, everything said about it), edit its triples, or Show on an artifact…, where it is drawn as a gold crown ♛. Classes stay out of instance counts, checks and reports, and are found by the search box.
- Several instances: shared values, Show path (shortest chain of shown links), One issue for these…. Two instances: Align… and Merge… (below); Between them lists every triple linking them (owl:sameAs too) with Issue…, Correct, Select link, Change… (predicate or direction) and Remove…, plus Suggest a new relation… (pick direction and predicate, then propose it — it waits in Issues to be accepted — or apply it now after a confirmation tick).
- Chosen links: flag each as an issue, mark them correct, or make one issue with the chosen instances.
- Long hover (≈ 1 s) shows the name and thickens all its lines.
- Every sheet is a view. The sheet bar lists all artifacts by group; opening one makes it the view's sheet. There is no main drawing.
- Documents with several pages (a PDF report or functional description) have one tab and a page bar: 1 page (‹ › or the page list), 12 at a time in a 4 × 3 grid — Next / Previous move one row of four and always keep twelve pages — or All pages in rows of four. Every page is a sheet of its own (anchor on it like any drawing); in the grid the other pages are windows showing what is placed on them, and ↗ works on that page. Pages render (≤ 4000 px) when first shown; the PDF is stored once in the session and project file.
- Side diagrams sit around the view's sheet on a grid, scaled to its size — added clockwise, from + bands next to any diagram, or from the chip row; drag one by its title to another slot; collapse, swap, open as its own view, or close. Each view remembers its own board; Reset folds them.
- Scope: This drawing (side diagrams dimmed until a chosen instance links into them), + Neighbours, Full graph.
- owl:sameAs twins are marked with the chosen instance and joined by a teal chain; click the chain (or The link… in the card's Same as box) to choose both and see, flag, change or remove the sameAs triple.
- Doors lead to sheets that are not on the board; guests are instances from another sheet dragged into a view without becoming a placement there.
- Evidence (top bar) shows, per chosen instance, a cropped call-out from every sheet it is on, its point-list rows, and where it is missing.
- Instances (top bar, or the link under Scope): every instance with one column per sheet (● shown there — click to go to it; – marked as not on it), point-list rows, the total, and Likely on (the sheet most of its neighbours are on).
- Filters: On no artifact (default), On no sheet, On a sheet, All; text and kind; sorting; CSV export.
- For chosen rows: place by hand (pick the sheet, then click), suggest placements with Claude on the sheets you tick (one call per sheet, only for the chosen instances; also for ones marked "not on it" before), mark them as not on a sheet, choose them in the view, or flag them. All its triples opens a card; closing it returns to the table.
- Rows are matched to instances when a cell equals a tag, label or name (ignoring case, spaces,
-,_,.). Header detection: the first line is a header unless it looks like data (e.g. a list of sensor names); First line is a header corrects it. - The list viewer: the table on the left (click, Ctrl/⌘- or Shift-click, or tick rows), the options for the chosen rows on the right — link to an instance (search, suggestions from the row, or the instance chosen in the app), unlink a wrong match (it stays rejected), report an issue (a flag on the instance, one group issue for several, or a note when no instance is linked), ignore rows. Rows that look like comments or noise (
####,# …,----) are found for you. Filters for Linked / Not linked / Reported / Ignored; undo and redo in the viewer. - Time series: a point (
brick:hasPoint/isPointOf) shows itsref:hasExternalReference→ref:TimeseriesReference/ref:hasTimeseriesIdunder Connections. Without one, pick a value from its point-list row (or type it) and Add: it inserts<point> ref:hasExternalReference [ a ref:TimeseriesReference ; ref:hasTimeseriesId "…" ]as one change.
- The reports cover the whole project and are the same on every sheet: not placed on any sheet yet; low-confidence placements (< 0.6); long flow links; flow asserted in both directions; anchors outside the drawing; open-ended segments; plus the review rules and any class or predicate you add as a report. Items on another sheet name it; clicking one goes there. Only what this view shows narrows the list to the current view.
- Reports are split into Instances, Links (predicates) and Mixed (group issues).
- One decision per report and item, stored once for the project: an instance can be Correct for one report and an Issue for another, and a decision made on one sheet holds on all. Issue / Unsure also flags the instance or link in Issues, naming the report. Click a chosen decision again to make it undecided.
- Review rules… say what the checks expect of which instances (has flow, is sensed, has a location, belongs to a system, is part of something, is typed, has a label, is placed), with exceptions — e.g. sensors and properties carry no flow.
- Classes & predicates… adds every instance of a class or every link of a predicate as a report; Mark all and decide together — a flag on each or one issue for all.
- Cards show only their header until chosen; choosing one highlights its instances and links. A red/green preview shows what a proposal or change does: removed links red, added links green, class and literal changes as chips, deletions and new instances ringed, moved anchors as arrows.
- Rules & general notes, flagged instances, links & group issues (each with the reports that raised it and an editable comment), proposed fixes (from Claude, from re-checks and from the reviewer) with accept / dismiss, then the change log and closed issues.
- Resolve with Claude: the items with editable comments (saved on the issue), bundles of related items with their own comment, and a comment on everything at the bottom. Fix types: re-point, remove, reverse and add links; add, delete, rename instances; change class; edit labels and comments; move nodes. When Claude proposes nothing it says why, and the answer stays on the issue.
- Rename an instance, edit its labels and comments, or any triple in and out of it (Edit… in its card). The name has a namespace (prefix) picker — any declared prefix or another namespace — beside the local name. The dialog shows what will change, warns about critical links (⚠: class, flow, points, location, parts, systems, connections, owl:sameAs, rdfs:subClassOf, time series, or an instance left without links or a label) and needs a confirmation tick. A rename rewrites every triple that uses the IRI — subject, object or predicate, nested values included — and moves its placements on every sheet, review records, issues and proposals along. A name that is already taken offers Merge into it….
- Triples in and out (in Edit…): every triple with the instance as subject or object, editable (predicate, object or subject), removable and addable. Each blank node is a dashed box headed by the triple that leads to it (
predicate → ◌ blank node · n triples inside), with its triples inside and nested blank nodes as nested boxes. - Merge two instances (Merge with… on a card, Merge… with two chosen): choose which IRI stays; identical triples become one, values both have are chosen (keep one, the other, or both), what pointed at the removed instance points at the kept one, links between the two are dropped; placements, review, issues and point-list matches move over (the kept one's own placements win). One change-log entry; Undo covers every artifact.
- Align two instances (Align with… / Align…): what A says compared with B — same, missing, differs, or parallel (B reaches the same thing through another predicate, in either direction, marked when it is the same kind of link). Per row: Leave, Add, or Use this instead; identifier-like values are marked as probably specific to one instance. Nothing is copied unless chosen; ⇄ swaps the direction.
- Relations: add, change or remove a triple between two chosen instances (above), as a proposal or straight into the change log.
- The loaded files are never modified. The graph on screen is the source graph with the active change-log entries applied in order (renames rewrite the IRI everywhere, also in earlier changes).
- Each entry records the triples removed (−) and added (+), the reason, the reviewer's comment, the origin, the time and notes. Revert takes an entry back out; it stays in the log as reverted.
- Exports:
- Refined graph (.ttl): the effective graph in one Turtle file, re-parsed before saving.
- Everything (.zip): at the top only the refined
.ttl, the project file (.project.json— open it to carry on) and a README;graph/original/(the graph files as loaded) andgraph/….corrections.sparql(the SPARQL Update that turns the original into the refined graph);review/(change log, review report, layout + review JSON, instance coverage);queries/(your saved SPARQL queries as.rqfiles andqueries.json);anchoring/(the anchoring prompt for every sheet and the record of runs);artifacts/(every drawing, document, image and point list, by group, withartifacts.csv). All artifacts (.zip) packs only the artifacts. - SPARQL Update: DELETE DATA / INSERT DATA, plus DELETE/INSERT … WHERE for renames and for edits inside blank nodes (found by where they hang and what they hold), and inserted blank nodes.
- Change log (.csv). Every CSV the app writes (change log, review report, instance coverage, SPARQL results) uses
;between columns. - Clean up leftovers…: placements, review marks, issues and point-list matches of entities no longer in the graph, set-aside or no-op change-log entries, and old
kept-from-…graph files — tick what to remove (Undo puts it back).
- Squares (4 × a typical symbol, 2–8 ×) or rectangles (corner to corner); up to 12 areas per call. Claude receives an overview, a clean and a tagged close-up per area, the reference images with notes, and the graph content inside each area. The areas are related asks Claude to compare them.
- Queries run in the browser on the effective graph with a small SPARQL 1.1 engine written for the app (
src/sparql.js): SELECT, ASK, CONSTRUCT, DESCRIBE, property paths, OPTIONAL, UNION, MINUS, FILTER, BIND, VALUES, sub-queries, aggregates, ORDER BY, LIMIT/OFFSET and the usual functions. - Ask in plain words: describe what you want; Claude writes the query from the graph's prefixes, classes and predicates, puts it in the editor and runs it — or refines the query already there.
- owl:sameAs as identity (on by default, OWL 2 RL equality rules). Prefixes in this project and Classes & predicates (with counts, sorted by count, prefix or name) insert into the query. Results list, expand to a big table (filter, sort, CSV) and highlight in the view.
- Saved queries: Save… stores the query in the editor with a name (from its first
# commentline, or the example's name) and an optional note; saving a query you opened from Saved offers Update or Save as new. Saved → Manage saved queries… runs, opens, renames, annotates and deletes them, imports.rq/.sparql/queries.jsonfiles and downloads them all as a zip. They are part of the project: the project file carries them, and the everything zip has one.rqfile per query (name and note as comments at the top) plusqueries.json.
- The sliders icon by the zoom buttons: zoom on click, zoom on Ctrl/⌘-click add, box select, links in the box, arrow direction (subject → object as written — the default — or the canonical direction: flow, container → part), and the display options.
| Setting | Default | Where |
|---|---|---|
| Scope | This drawing | Side panel → Scope |
| Connection categories | Flow on #2563eb · Sensing & points on #16a34a · Location on #d97706 · System topology off #7c3aed · Parts & composition on #db2777 · Heat transfer on #dc2626 · Physical connection on #0f766e · Other links off #64748b |
Side panel → Connections (per category and predicate; colours editable) |
| Merge parallel predicates into one line | on | Connections / View settings |
| Fade instances without a shown connection | on | Connections |
| Show class under names | on (T toggles) | Display / View settings |
| Node name | Instance name (IRI local part) | Display (or rdfs:label) |
| Label size · node size · line width | 11 px · 3 × · 2 × | Display |
| Hide labels that would overlap | on | Display / View settings |
| Zoom to an instance when clicked | on | View settings |
| Zoom when adding with Ctrl/⌘-click | off | View settings |
| Box select by plain drag | off (Shift-drag always) | Zoom bar / View settings |
| Box also picks links | off | View settings |
| Arrows | subject → object as written | View settings |
| Only what this view shows (reports) | off | Review & checks |
| Low-confidence threshold | 0.6 (dotted amber ring) | Fixed |
| Drawing ⇄ Graph fader | centre (both 100 %) | Bottom of the canvas |
| Re-check areas | square, 4 × symbol, max 12 | Re-check bar |
| SPARQL | Show results in the view: on · owl:sameAs: on | SPARQL panel |
| Theme | Auto | Top bar |
| Claude model inside Claude | claude-opus-5-5, tier complex |
Fixed |
| Claude model with an API key | claude-opus-5-5, max 32 000 output tokens |
Claude API key dialog or .env |
| Undo history | 200 steps | Fixed |
Kept per browser in localStorage: kgov:ds:<sheet id> (the project record: placements of the first sheet, review, per-report decisions, rules, findings, change log, point-list edits), kgov:sheet:<sheet id> (placements of further sheets), kgov:board (each view's side diagrams), kgov:settings, kgov:theme, kgov:checkView, kgov:reviewPicks, kgov:sparql*, and kgov:api (only when you tick Remember). The last session (graphs, artifacts, groups, notes, point lists, description) is kept in IndexedDB kg-overlay-inspector / last.
Published artifact: runtime contract 0.2.85; capabilities sample (with images) and downloads; sharing is set from the artifact's Share menu.
- Line style: solid = component ↔ component; dashed = one end is a room, zone, storey or system.
- Arrows point from subject to object of the asserted triple (or, by choice, in the canonical direction). Parallel predicates share one line.
- Highlights: the issue in question is orange with a glow and rings, and the rest dims; highlights breathe gently (not with reduced motion). Chosen links get a blue halo.
- Node shapes: ● component, ○ duct/pipe segment, ■ fitting, ▢ location, ⬡ system, ◆ point/sensor. A dotted ring means low placement confidence.
- Hidden from the canvas:
rdf:type(shown under names),rdfs:*,owl:*,skos:*,dcterms:*and literals — all in the instance card.
- Equipment, dampers, terminals and sensors: centre of the symbol. Motors and actuators: their own symbol.
- Ducts and pipes: midpoint of the drawn run between their two flow neighbours. Junctions: where the lines meet.
- Rooms and zones: their name/number text. Systems: a system label or a representative line point.
- Catalog and product pages: the row or figure that shows the item.
- Not drawn on a sheet: reported as not on it.
| Field | Content |
|---|---|
artifact, units |
the first sheet, px or norm1000 |
anchors, offDrawing, guests |
placements of the first sheet |
qa |
{nodes, edges, checks} — review records; checks holds one decision per report and item (`" |
findings, recheckRuns, rules, checkRules, changes |
the review state (rules include group issues; findings include reviewer proposals) |
sheets |
further sheets: {id, role, title, group, name, width, height, page, anchors, offDrawing, runs, guests} |
{id, seq, batch, at, status: applied|reverted, type, title, why, comment, origin, conf, del:[triple], ins:[triple], notes, anchor, resolved, rename?:{from, to}}
A triple is {s, p, o, ot: iri|lit|blank|bnode, lang, dt}; inserted blank nodes use ot: 'bnode' and subjects _:id.
graphs [{name, text}] (unmodified sources), drawing and sheets (role, title, group, comment, layout, dataUrl), images (role legend|title|context, group, comment, dataUrl), tables [{name, text, group, comment, links, unlinks, ignore, flags, headerMode}], description, layout. The API key is never included.
- Review report:
item, name, why_in_report, iri_or_link, classes, status, note, placement, x, y, confidence, reviewed_at. - Change log:
change, status, applied_at, reverted_at, type, title, removed_triples, added_triples, why, reviewer_comment, origin, confidence, finding, batch, notes. - Instance coverage:
instance, iri, label, class, kind, <one column per sheet>, point list rows, total. - SPARQL results: one column per variable.
src/head.html styles (light/dark tokens) and <title>
src/body.html markup
src/app.js all logic
src/sparql.js the SPARQL 1.1 engine (also loadable in Node)
src/utif.js UTIF.js TIFF decoder by Photopea (MIT, github.com/photopea/UTIF.js), bundled for offline TIFF scans
n3bundle.js N3.js 2.0.1 Turtle/TriG parser, inlined (bundle_n3.mjs rebuilds it)
assemble.py → site/index.html (published page) and site/test.html (standalone file)
tools/make_demo_sheets.py draws the made-up demo artifacts
site/demo/ demo graph, drawing, legend, layout and the made-up extra artifacts (sheets.json)
tests/ Playwright tests (t2–t45), mock_claude.js (stand-in for Claude), fixtures,
sparql_engine.js / sparql_sameas.js (Node tests for the engine)
Build and test:
python3 assemble.py
cp site/test.html graph-overlay-inspector.html # the single-file copy
cd site && python3 -m http.server 8765 & # tests expect this server
cd .. && python3 tests/t25_review.py tests/out tests tests/fixtures
node tests/sparql_engine.js "$PWD" site/demo/HTRK-VEN03_knowledge_graph.ttl
node tests/sparql_sameas.js "$PWD"
Tests need Python Playwright with Chromium. They run against the stand-in for Claude and a mocked API, so no real key is used; the .env fixtures they need hold fake keys and are written at test time (never committed).
- Claude features are tested with stand-ins. How good the real placements and proposals are only shows on real runs.
.ttlcan't be saved directly inside Claude (the page's allowed file types); use the single HTML file, the package,.txtor Copy.- The refined Turtle is one merged file; several loaded graphs are not split back into their sources (the package keeps the sources and the SPARQL Update).
- Projects move between browsers or people only via the project file.
- The demo's extra artifacts are made up (AHU-01 detail sheet, BMS screenshot, catalog page, point list, sensors); the HTRK-VEN03 overview and graph are the real ones.
- owl:sameAs reasoning is query-time only (SPARQL); the drawing shows twins but does not merge them.
- The API key route only works outside claude.ai, and the key is as safe as the browser profile it is remembered in.