Knowledge Intelligence for Rapid Analysis
An industrial AI operating system — plant documents and live operational data in, decisions out.
Next.js 15 · FastAPI · PostgreSQL · Neo4j · Qdrant · Redis · MinIO · Docker
Industrial plants run on knowledge that's scattered and stale. Decades of manuals, P&IDs, inspection reports, maintenance logs, and compliance records live in disconnected PDFs, spreadsheets, and filing cabinets. When something fails at 2am:
- The engineer who knew the fix retired years ago
- The relevant SOP is buried in a 400-page PDF nobody can search
- Maintenance history, compliance status, and equipment data live in three different systems that don't talk
- Answers come from tribal memory, not evidence — and mistakes are expensive and unsafe
The data exists. It's just fragmented, unsearchable, and disconnected from live plant state.
Turn every document and every live signal into one connected, queryable knowledge layer — and put an AI on top that always cites its source.
- Ingest everything — PDFs, spreadsheets, images, CSVs go through auto text + entity extraction
- Connect it — extracted entities (equipment, standards, failures) become a knowledge graph, fused with live plant state: work orders, failures, technicians, open violations
- Ground every answer — the Copilot orchestrates retrieval + graph + analytics tools; every response is traceable to the exact source passage, with a confidence score
- Close the loop — compliance violations detect automatically, follow maintenance progress, and auto-close when the condition clears — no manual reconciliation
| Role | Value |
|---|---|
| Plant engineer | Ask a question, get a cited answer in seconds instead of digging through PDFs |
| Maintenance lead | One board for work orders, MTTR/cost analytics, and auto-assignment — no spreadsheet juggling |
| Compliance officer | Violations surface and close themselves as equipment is repaired; audit trail is automatic |
| Operations | Live plant health, availability, and risk on one dashboard — decisions from evidence, not memory |
Try the app directly — no setup required:
| Live site | straining-crumpet-handset.ngrok-free.dev |
| Demo account | demo@kira.example.com / password-password123 |
| Simulation | straining-crumpet-handset.ngrok-free.dev/sim-control.html — drive a live plant simulation |
Prereqs: Docker Desktop + Node 20+
cp .env.example .env
npm run infra:up # builds & starts all 8 services| Service | URL |
|---|---|
| Web app | http://localhost:3000 |
| API docs (OpenAPI) | http://localhost:8000/docs |
| Neo4j browser | http://localhost:7474 |
| MinIO console | http://localhost:9001 |
Create a workspace on the web app, upload a document, ask the Copilot. Runs fully offline — no LLM key required (set LLM_PROVIDER + LLM_API_KEY in .env to upgrade answers).
| 📄 Documents | PDF/DOCX/XLSX/CSV/images → auto-extract → entities → graph. Public/private visibility by role. |
| 🕸️ Knowledge Graph | 3D force graph: equipment, standards, failures, work orders, technicians, violations — docs + live plant state |
| 🔧 Maintenance | Work-order kanban with a validated state machine, engineer assignment, MTTR/cost analytics |
| 🛡️ Compliance | Rules engine detects violations, mirrors maintenance progress, auto-closes when the condition clears |
| 📊 Analytics | Plant health, availability, compliance score, AI quality metrics, CSV export |
| 🤖 KIRA Copilot | Streaming chat with citations + confidence, orchestrating retrieval/graph/analytics/compliance tools |
| 🎛️ Simulation | localhost:3000/sim-control.html — drive a live plant simulation (degrade → fail → repair) |
Lead creates the org and approves engineers. Engineers register with the same org name → join request → lead approves from the bell or Team page. All data is org-scoped.
# backend tests (~100, no containers needed)
cd apps/backend && python -m pytest tests/
# frontend typecheck
cd apps/web && npx tsc --noEmit
# rebuild after changes
docker compose -f docker/docker-compose.yml up -d --build backend web
# seed demo data
python samples/seed_equipment.py # 24 assets
python samples/simulate.py # live plant simulation (Ctrl+C to stop)
bash samples/seed.sh # sample CSVs → documents → graph- Register on
/register— the first user for an org name becomes its Lead. Later users registering with the same org name become Engineers in a pending state until the lead approves them (bell → Approve, or the Team page). - Seed data (optional):
python samples/seed_equipment.pyloads ~24 assets;bash samples/seed.shuploads sample CSVs and builds the graph. - Upload a document → open it → Analyze to extract entities into the graph.
- Knowledge Graph → Rebuild to project documents + live plant state (work orders, failures, technicians, violations).
- Compliance → Run Evaluation to detect violations; Maintenance to raise and work orders.
- Simulate a living plant:
python samples/simulate.py(or open/sim-control.html) degrades assets, opens work orders, and repairs them on a loop.
equipment goes offline ──► compliance detects a violation
│ │
▼ ▼
work order raised ──► engineer assigned (least-loaded)
│ │
▼ ▼
WO completed ──► asset restored ──► violation auto-closes
│
▼
graph + analytics + dashboard all reflect it on next refresh
Nothing is reconciled by hand — repairing an asset closes its compliance findings, and the knowledge graph reflects live state on every Rebuild.
Everything is env-driven (.env, see .env.example). Sensible local defaults ship in the compose file.
| Group | Keys | Default |
|---|---|---|
| Auth | JWT_SECRET_KEY, ACCESS_TOKEN_EXPIRE_MINUTES, REFRESH_TOKEN_EXPIRE_DAYS |
30 min / 7 days |
| Postgres | DATABASE_URL / POSTGRES_* |
local compose |
| Neo4j | NEO4J_URI, NEO4J_USER, NEO4J_PASSWORD |
graph backend |
| Qdrant | QDRANT_URL, QDRANT_COLLECTION |
kira_chunks |
| MinIO | MINIO_ENDPOINT, MINIO_BUCKET, MINIO_SECURE |
object store |
| AI | LLM_PROVIDER, LLM_MODEL, LLM_API_KEY, EMBEDDING_DIM |
none / 384-dim |
LLM_PROVIDER=none(default) runs fully offline: heuristic entity extraction, deterministic summaries, hash-based embeddings. Set a provider + key to upgrade answers without any code change.
- JWT access + rotating refresh tokens; refresh rotation detects reuse
- Server-side sessions — revoke any session instantly from Settings
- Account lockout after repeated failed logins
- DB-backed RBAC — 30+ granular
<resource>.<action>permissions (e.g.maintenance.approve,documents.delete); every API route re-checks, the UI only hides what you can't reach - Strict multi-tenancy — every query is org-scoped; no cross-org read is possible
cd apps/backend && python -m pytest tests/ # 19 suites: auth, RBAC, documents,
# maintenance, compliance, graph, analytics, copilot
cd apps/web && npx tsc --noEmit # frontend typesTests use in-memory SQLite + in-memory storage/graph backends — no containers required.
| Doc | Covers |
|---|---|
| Project Spec | Vision, modules, roles, tenancy |
| Architecture | Services, data flows, deployment |
| Backend Spec | Layering, state machines, testing |
| Frontend Spec | Pages, conventions, design tokens |
| API Spec | Every endpoint group |
| Database | Tables, relationships, migrations |
| AI Agents | Orchestrator, tools, offline-first |