Dashboard
● Checking… KB: — ✦ Aria
✦
Hi, I'm Aria — your AI Receptionist
I use Retrieval-Augmented Generation (RAG) + GPT-4o-mini to predict appointment delays before they happen, rank risk, and draft reminders — all with human approval. Built by Vic Parmar · GGU Doctoral Research · 2026
KB Records
—
Historical appointments
This Week
—
Upcoming appointments
High Risk
—
Risk score ≥ 70
Low Risk
—
Risk score < 45
🗓 Upcoming Appointments at a Glance
IDDateTimeTypeLocation BufferPunct.Priority
Loading…
🔬 Quick Risk Snapshot
📊
Click Analyze Now to run the RAG pipeline and see risk scores.
⚡ Full Analysis Pipeline
RAG + GPT-4o-mini

Run the full three-step pipeline: (1) hybrid rule-based retrieval scoring, (2) GPT-4o-mini narrative generation grounded in KB evidence, (3) approval-gated reminder drafts for high-risk appointments.

1 Rule-Based Scoring
›
2 LLM Narratives
›
3 Draft Reminders
Ready. Click "Analyze Risk" to start.
📅 Upcoming Appointments
IDDateTimeTypeLocation BufferPriorityPunct. Score
Loading appointments…
📚 Knowledge Base Records

Historical appointment records with outcomes and failure tags. Used for RAG retrieval.

IDDateTypeLocation OutcomeDelayFailure TagsPunct.
Loading KB records…
ℹ System Architecture

LLM Model

GPT-4o-mini (OpenAI API) · Temperature 0.2 · Evidence-grounded · Citation-enforced

Retrieval Strategy

Hybrid BM25 + Jaccard similarity · Temporal weighting (90-day × 1.5) · Metadata pre-filtering

Embedding Model

text-embedding-3-small (OpenAI) · 1536-d vectors · ~$0.00002/1K tokens

Vector Store

Azure AI Search (hybrid BM25 + vector) · RRF merging · Enterprise-grade access control

RAG Enhancements

8 techniques: multi-query expansion, temporal decay, metadata filtering, RRF, re-ranking, evidence thresholding, CoT, self-consistency

Safety & Governance

All drafts approval-gated · Human-in-loop enforced · No auto-send · PII pseudonymised · PDPA/GDPR compliant

Performance

Rule-based p50: ~120ms · LLM narrative p50: ~1.6s · Token cost: ~$0.0003/appointment · Precision@3 = 1.00

Infrastructure

Azure App Service B1 (always-on) · Linux · Node 20 LTS · Southeast Asia (Singapore) · aria.vicparmar.com

🔗 API Endpoints
MethodEndpointDescription
GET/healthHealth check, KB size, upcoming count
GET/list-upcomingAll upcoming appointments as JSON
GET/list-kbAll KB history records as JSON
POST/analyze-weekRun rule-based risk scoring
POST/analyze-with-llmRule-based + GPT-4o-mini narratives
POST/draft-remindersGenerate approval-gated reminder drafts
POST/appointmentsAdd new upcoming appointment
POST/kbRecord outcome into KB (feedback loop)
📖 Academic References
  1. Anthropic (2026). The Labor Market Impact of AI: An Analysis of Task-Level Exposure. anthropic.com/research/labor-market-impacts — found 14% slower hiring for workers aged 22-25 and a 33% coverage gap between benchmark and observed AI task exposure.
  2. Lewis, P. et al. (2020). Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks. NeurIPS 2020. — foundational RAG paper.
  3. Wei, J. et al. (2022). Chain-of-Thought Prompting Elicits Reasoning in Large Language Models. NeurIPS 2022.
  4. Robertson, S. & Zaragoza, H. (2009). The Probabilistic Relevance Framework: BM25 and Beyond. Foundations and Trends in IR.
  5. Wang, X. et al. (2022). Self-Consistency Improves Chain of Thought Reasoning in Language Models. ICLR 2023.
  6. OpenAI (2024). GPT-4o Technical Report. openai.com.