AuRAG connects plant equipment, failures, work orders, procedures, people, documents, and regulatory clauses in a Neo4j knowledge graph. A LangGraph supervisor routes operator questions to Copilot, RCA, Compliance, or Lessons Learned agents, returning cited answers and an inspectable graph trail.
The current product also includes continuous ingestion, proactive telemetry, persistent work-order decisions, cross-session mem0 memory, GraphRAG versus dense-only comparison, knowledge-retirement risk, and a durable RAGAS evaluation dashboard.
See PRD.md, PRD_CLOSURE.md, and docs/ACCEPTANCE_REPORT.md.
frontend/: Next.js 16 operator console with separate workspaces.backend/: FastAPI routes for chat, graph, telemetry, evaluations, comparison, ingestion, knowledge risk, notifications, and work orders.agents/: LangGraph supervisor and four specialist agents.retrieval/: Neo4j vector, Qdrant dense, BM25, graph traversal, and configurable local/Cohere reranking.ingestion/: clean text, Tesseract plus cloud OCR, Gemini P&ID extraction, object storage, Watchdog, Redis, and RQ.telemetry/: synthetic signal generation, failure-pattern matching, and autonomous predictive-event creation.evaluation/: hard-gated RAGAS scoring and GraphRAG comparison validation.infra/: Neo4j schema/seed and persistent local Compose services.
Copy .env.example to .env. Required for the complete local product:
NEO4J_URI
NEO4J_USERNAME
NEO4J_PASSWORD
REDIS_URL
QDRANT_URL
GROQ_API_KEY
GEMINI_API_KEY
MEM0_API_KEY
Provider-specific options:
RERANK_PROVIDER=local
COHERE_API_KEY=
CLOUD_OCR_PROVIDER=google_vision
GOOGLE_VISION_API_KEY=
The local reranker is the default and requires no Cohere key. Google Cloud
Vision is implemented as the PRD-named low-confidence OCR fallback; auto
mode uses it when GOOGLE_VISION_API_KEY is present and otherwise retains
the Gemini compatibility fallback.
RAGAS uses llama-3.3-70b-versatile by default and fails the acceptance gate
on provider errors, missing scores, or any metric below 0.70.
GROQ_REASONING_MODEL and GROQ_ROUTING_MODEL can override the agent models
when a provider model is temporarily quota-limited.
From the repository root:
powershell.exe -NoProfile -ExecutionPolicy Bypass -File .\scripts\bootstrap.ps1
powershell.exe -NoProfile -ExecutionPolicy Bypass -File .\scripts\bootstrap.ps1 -SkipInstall -SkipInfra -InitializeDataStart the API:
.\.venv\Scripts\python.exe -m uvicorn backend.app.main:app --host 127.0.0.1 --port 8000Start the frontend:
npm.cmd --prefix frontend run devOptional workers:
.\.venv\Scripts\python.exe -m telemetry.worker
.\.venv\Scripts\python.exe -m ingestion.workers.watcher
.\.venv\Scripts\rq.exe worker aurag-ingestFor managed deployments, documents should be uploaded through
POST /api/ingestion/documents; the API writes an immutable object to shared
S3-compatible storage and enqueues the same RQ ingestion pipeline.
Backend and domain tests:
$env:PYTEST_DISABLE_PLUGIN_AUTOLOAD='1'
.\.venv\Scripts\python.exe -m pytest -p no:cacheproviderFrontend:
npm.cmd --prefix frontend test -- --run
npm.cmd --prefix frontend run lint
npm.cmd --prefix frontend run build
npm.cmd --prefix frontend run test:e2e
npm.cmd --prefix frontend run test:e2e:live
npm.cmd --prefix frontend audit --omit=devThe normal Playwright suite uses deterministic API fixtures for UI behavior.
test:e2e:live is a separate non-mocked gate that submits a real question to
FastAPI, renders graph evidence, and loads real mobile knowledge-risk data.
Provider-backed acceptance:
docker build -t aurag:acceptance .
docker run --rm --env-file .env `
-e NEO4J_URI=bolt://host.docker.internal:7687 `
-e NEO4J_USERNAME=neo4j `
-e NEO4J_PASSWORD=aurag-local-password `
-e QDRANT_URL=http://host.docker.internal:6333 `
-e REDIS_URL=redis://host.docker.internal:6379/0 `
-v aurag_acceptance_hf_cache:/home/aurag/.cache/huggingface `
aurag:acceptance python -m evaluation.validate_ragasLive mem0 proof:
$proofId = [guid]::NewGuid().ToString('N')
docker run --rm --env-file .env `
-e PROOF_ID=$proofId `
-e PYTHONPATH=/app `
-v "${PWD}/scripts/validate_mem0.py:/tmp/validate_mem0.py:ro" `
aurag:acceptance python /tmp/validate_mem0.pyBackend smoke:
powershell.exe -NoProfile -ExecutionPolicy Bypass -File .\scripts\smoke.ps1render.yamldefines the API, telemetry worker, and RQ ingestion worker.frontend/vercel.jsondefines the frontend build.BACKEND_CORS_ORIGINSmust contain the exact deployed frontend origin.- Render/Vercel URLs and cloud smoke evidence require access to those accounts; configuration files alone are not deployment proof.
See docs/DEPLOYMENT.md.
As of July 25, 2026:
- No Render or Vercel account access is available, so hosted URLs, deployed browser acceptance, deck upload, and demo video are not proven.
- Cohere reranking is live-validated.
- Google Vision OCR reaches the provider, but its configured Google Cloud project has billing disabled and returns HTTP 403.
- The public 42-page P&ID course document uses tags unrelated to the
synthetic plant. Two manually reviewed diagram pages are provider-validated;
it is not valid evidence for plant-specific
CONNECTED_TOrelationships.