analytics

AI Quality Evaluation

Gemma 4 triaj doğruluk metrikleri & RAG kalite ölçümü

gemma4:e4b | Ollama local | 2026-05-11 | No cloud API

wifi

Live System Status

Yükleniyor...

—

Toplam Oturum

—

🔴 RED Triaj

—

Guardrail Escalations

—%

Avg Completeness

93%

Overall Score

14/15 tests passed

100%

Triage Accuracy

5/5 cases correct

100%

RAG Retrieval

6/6 sources correct

100%

JSON Validity

All outputs valid JSON

memory

AI Execution Environment

Model

gemma4:e4b

Runtime

Ollama local

External AI API

✗ None

RAG Backend

ChromaDB (local)

Embedding

multilingual-MiniLM (local)

Patient Data Egress

✗ None

Avg Latency

11–39s (GPU)

RAG Chunks

~90

emergency

Triage Decision Results

5/5 correct ✓

Case	Expected	Predicted	Confidence	Result

database

RAG Retrieval Accuracy

6/6 correct ✓

Query	Expected Source	Retrieved	Relevance	Result

Fix Note (2026-05-11): Prior to this date, cardiac chest-pain query retrieved ENT_Emergency instead of Cardiac_Emergency. A dedicated Cardiac_Emergency chunk was added with explicit STEMI/ACS keyword anchors — resolving the retrieval miss.

shield

Safety Guardrail Layer

Deterministic Rules

Gemma 4 produces the structured clinical assessment, but deterministic safety rules can escalate high-risk findings — ensuring life-threatening presentations are never under-triaged, even if the LLM produces an incorrect output.

🔴 AUTO-ESCALATE TO RED

• Chest pain + left arm radiation + diaphoresis
• SpO₂ <90% (vital sign threshold)
• Systolic BP <90 mmHg (shock)
• "Worst headache of my life" (SAH)
• Facial droop / arm weakness / speech (FAST stroke)
• Anaphylaxis pattern (throat swelling + allergy)
• GI haemorrhage (bloody stool / haematemesis)
• Fever + petechiae/purpura (meningococcal sepsis)

🟡 AUTO-ESCALATE TO YELLOW

• Infant / young child with fever
• Rigors + high fever (sepsis pattern)
• Pulse >130 or <40 bpm (vital sign threshold)
• RR >30 or <8 /min (respiratory failure)
• Temperature ≥41°C (hyperpyrexia)

Disclaimer: These are synthetic evaluation cases for development purposes. Real-world accuracy requires clinical validation with licensed healthcare professionals. AnamnezAI is not a diagnostic system — all outputs require physician review.
Generated locally by AnamnezAI using Gemma 4 via Ollama. No patient data transmitted externally.