📚 Complete Tutorial: Medical Chatbot Configuration with AI Engine
💰 TOKEN SYSTEM AND COST OPTIMIZATION
🔤 WHAT ARE TOKENS?
Tokens are the basic units used for billing AI services. One token represents approximately:
📝 1 token = 4 characters in English
📝 1 token = 1 short word in foreign language
📖 1000 tokens ≈ 750 words
Practical examples:
Simple question: “what medicines for cough?” → ~10 tokens
Short medical answer (200 words) → ~250 tokens
Complete conversation (question + answer) → ~300-500 tokens
🏪 OPENROUTER – THE AI “SUPERMARKET”
OpenRouter functions as a centralized marketplace where different AI providers compete on price:
YOUR APPLICATION → OPENROUTER → [Provider1, Provider2, Provider3...] ↓ Automatically chooses the lowest price
🎯 HOW IT WORKS:
A single API Key for all providers
Automatic selection of the cheapest available model
Unified billing – a single monthly invoice
Transparent switching between providers without code change
📊 TOKEN PRICE COMPARISON TABLE
| Provider | Model | Input (1M tokens) | Output (1M tokens) | Performance |
|---|---|---|---|---|
| OpenRouter | DeepSeek | $0.14 | $0.28 | ✅ Excellent |
| OpenRouter | Mistral 7B | $0.14 | $0.42 | ✅ Good |
| OpenRouter | Llama 3.1 8B | $0.18 | $0.54 | ✅ Very Good |
| OpenAI Direct | GPT-3.5 Turbo | $0.50 | $1.50 | ✅ Good |
| OpenRouter | Claude Haiku | $0.25 | $1.25 | ✅ Excellent |
| OpenAI Direct | GPT-4 Turbo | $10.00 | $30.00 | ✅ Exceptional |
| Google Direct | Gemini Pro | $0.50 | $1.50 | ✅ Good |
💡 CONVERSATION COST CALCULATION:
Typical Conversation: 500 tokens With DeepSeek: (500 ÷ 1,000,000) × $0.28 = $0.00014 With GPT-4 Turbo: (500 ÷ 1,000,000) × $30.00 = $0.01500
🏆 COST-PERFORMANCE RECOMMENDATIONS:
🥇 DEEPSEEK (via OpenRouter)
Best price-quality ratio
Cost: 10x lower than GPT-4 Turbo
Performance: Excellent
🥈 MISTRAL (via OpenRouter)
Cost similar to DeepSeek
Specialized for European languages
🥉 GPT-3.5 TURBO (direct)
Good for prototyping
Decent cost, moderate performance
🚀 COST OPTIMIZATION STRATEGY
1. FOR TESTING/DEVELOPMENT:
OpenRouter + DeepSeek: $0.14/M input - Minimum cost with good performance - Ideal for fast iterations
2. FOR QUALITY PRODUCTION:
OpenRouter with fallback: DeepSeek → Mistral → Claude - Perfect cost-quality balance - Automatic redundancy
3. FOR MAXIMUM ACCURACY:
GPT-4 Turbo direct: $10.00/M input - Most obedient to prompts - Guaranteed quality, higher cost
👉 OpenRouter offers the best cost-flexibility ratio for medical projects!
🗃️ MEDICINE DATABASE (JSON)
[
{
“name”: “Amoxicillin”,
“category”: “Antibiotic”,
“prescription”: “Yes”,
“indications”: “Bacterial respiratory infections, Dental infections, Urinary tract infections. Administered only on a doctor’s recommendation.”
},
{
“name”: “Ciprofloxacin”,
“category”: “Fluoroquinolone Antibiotic”,
“prescription”: “Yes”,
“indications”: “Urinary tract infections, Gastrointestinal infections, and Respiratory infections. Requires a medical prescription.”
},
{
“name”: “Metronidazole”,
“category”: “Antibiotic / Antiparasitic”,
“prescription”: “Yes”,
“indications”: “Anaerobic infections and Parasitic infections. Administration only under medical supervision.”
},
{
“name”: “Prednisone”,
“category”: “Corticosteroid”,
“prescription”: “Yes”,
“indications”: “Inflammatory and autoimmune conditions. Administered only upon the doctor’s indication.”
},
{
“name”: “Enalapril”,
“category”: “Antihypertensive (ACE Inhibitor)”,
“prescription”: “Yes”,
“indications”: “Arterial hypertension and heart failure. Requires medical monitoring.”
},
{
“name”: “Diazepam”,
“category”: “Anxiolytic / Sedative”,
“prescription”: “Yes”,
“indications”: “Anxiety disorders and insomnia. Uncontrolled administration can lead to dependence.”
},
{
“name”: “Insulin”,
“category”: “Antidiabetic Hormone”,
“prescription”: “Yes”,
“indications”: “Treatment of diabetes mellitus. Dosage is determined exclusively by the doctor.”
},
{
“name”: “Paracetamol”,
“category”: “Antipyretic / Analgesic”,
“prescription”: “No”,
“indications”: “Fever and mild pain. Available without prescription.”
},
{
“name”: “Ibuprofen”,
“category”: “Non-steroidal Anti-inflammatory Drug (NSAID)”,
“prescription”: “No”,
“indications”: “Muscle pain, Joint pain, Dental pain, and fever. No prescription required for low doses.”
},
{
“name”: “Loratadine”,
“category”: “Antihistamine”,
“prescription”: “No”,
“indications”: “Seasonal allergies and hives. Can be purchased without a prescription.”
}
]
🎯 PROJECT GOAL
Creating a medical chatbot that:
Answers EXCLUSIVELY from its own medicine database
Never invents medicines or indications
Provides short and precise answers (under 200 words)
Recommends medical consultation for treatment
Uses cost-optimized infrastructure
🏗️ SYSTEM ARCHITECTURE
Data Flow:
User → Chatbot AI Engine → Qdrant Vector DB → Medicines JSON → Response
Critical Components:
AI Engine Pro – the chatbot platform
Qdrant Cloud – vector database
OpenAI GPT-4 Turbo – optimized AI model
PDF/JSON – the medicine database
⚙️ CHATBOT CONFIGURATION
1. “CHATBOT” TAB – BASIC SETTINGS
📍 AI Engine → Chatbots → [Chatbot Name] → “Chatbot” Tab
📋 GENERAL INFORMATION:
Name: Zerodoping ID: chatbot-xxxxxxx (auto-generated) Scope: chatbot Local Memory: Yes (preserves conversation history)
🎯 OPERATING MODE:
Mode: Chat (✅ Selected) - Conversational - maintains discussion context - Ideal for long medical interactions
📝 SYSTEM INSTRUCTIONS (SYSTEM PROMPT) – COMPLETE:
You are a medical assistant that uses EXCLUSIVELY the medicine database. ABSOLUTE RULES: 1. You have access to ONLY ONE source of information: our medicine database 2. All answers must come ONLY from this database 3. You are not allowed to use any other medical knowledge from your training 4. If a medicine is not in our database, it does not exist for you 5. The database is your only source of truth When you receive a question: 1. Search our database 2. If you find something relevant: answer ONLY with what you found 3. If you do not find: "I do not have relevant medicines in our database"
🎯 ADVANCED MEDICAL PROMPT STRATEGY:
🏗️ ENGINEERING STRUCTURE:
✅ ROLE DEFINITION: "medical assistant" ✅ RADICAL RESTRICTION: "EXCLUSIVELY the database" ✅ EXPLICIT RULES: 5 absolute points + 3 action steps ✅ COMPLETE AMBIGUITY ELIMINATION: "it does not exist for you" ✅ UNIQUE SOURCE OF TRUTH: "your only source" ✅ CLEAR ALGORITHM: "When you receive a question: 1. Search... 2. If you find... 3. If you do not find..."
🔬 PSYCHOLOGICAL ANALYSIS OF THE PROMPT:
🧠 COGNITIVE CONSTRAINTS:
1. "EXCLUSIVELY" → Prohibits access to general knowledge 2. "ONLY from this database" → Eliminates alternative sources 3. "You are not allowed" → Limits freedom of action 4. "It does not exist for you" → Rewrites the AI's reality 5. "Your only source" → Defines a new ontology
⚡ ACTION MECHANISM:
INPUT: User question ↓ PROCESSING: "Search our database" ↓ DECISION: ├─ If found → "answer ONLY with what you found" └─ If NOT found → "I do not have relevant medicines" ↓ OUTPUT: 100% database-based answer
🎯 BENEFITS OF THE PROMPT ARCHITECTURE:
✅ MAXIMUM DETERMINISM:
Same input → always the same output
Eliminates unwanted variations
Predictable behavior
✅ MEDICAL SECURITY:
Zero “hallucinations” or inventions
Answers only from verified sources
Protection against misinformation
✅ OPTIMIZED PERFORMANCE:
Direct search in the database
No unnecessary processing of general knowledge
Fast and precise answers
🔧 ANALYSIS OF SETTINGS FROM THE SCREENSHOT:
✅ CORRECTLY CONFIGURED:
Scope: chatbot → optimal for user interaction
Local Memory: Yes → maintains medical context per session
Mode: Chat → ideal for medical consultations
Complete Prompt → covers all medical scenarios
🎯 IMPACT OF SETTINGS:
Memory active → the chatbot remembers the conversation history
Chatbot scope → complete conversational functionality
Exhaustive prompt → ensures 100% adherence to the database
2. “AI MODEL” TAB – MODEL AND PERFORMANCE CONFIGURATION
📍 AI Engine → Chatbots → [Chatbot Name] → “AI Model” Tab
🤖 PROVIDER AND MODEL SELECTION:
Environment: OpenAI Model: GPT-4 Turbo (✅ Selected)
🎯 MODEL SELECTION ANALYSIS:
🏆 GPT-4 TURBO – TECHNICAL MOTIVATION:
✅ Performance: Best quality-price ratio in the GPT-4 range ✅ Obedience: Exceptionally respects strict prompts ✅ Context: 128k tokens - enough for long medical history ✅ Speed: Optimized for fast answers ✅ Cost: ~3x cheaper than standard GPT-4
⚙️ ADVANCED TECHNICAL PARAMETERS:
🌡️ TEMPERATURE: 0.1
Value: 0.1 (very low) Impact: - Maximum deterministic answers - Minimizes variation and creativity - Same input → always the same output - Ideal for medical accuracy Motivation: Eliminates any possibility of "hallucination"
📊 MAX TOKENS: 500
Value: 500 tokens Calculation: - Typical Question: ~20 tokens - Medical Answer: ~150-300 tokens - Safety Buffer: ~180 tokens Benefit: - Prevents overly long answers - Cost control - Forces medical conciseness
🧠 CONTEXT CAPACITY:
Contextual: 128,000 tokens (disponibil) Completion: 4,096 tokens (limitat) Recommended: 4,096 tokens (recomandat) Strategie: - We use 4,096 for cost optimization - Enough for long medical conversations - Maintains relevant history without overhead
🚫 CRITICAL CONFIGURATION: FILE UPLOAD
Uploads: File Upload ❌ DISABLED Motivation: - Forces exclusive use of the Knowledge Base - Prevents loading external documents - Ensures the AI uses only our database Impact: The key that blocks "leaks" of general knowledge
🔧 ADVANCED CONFIGURATION ANALYSIS:
🎯 COST-PERFORMANCE STRATEGY:
Model: GPT-4 Turbo → $10/M input (vs $30/M standard GPT-4) Temperature: 0.1 → reduces unnecessary tokens Max Tokens: 500 → limits costs/output Context: 4,096 → optimized for medical conversations
🛡️ SAFETY MECHANISMS:
1. File Upload disabled → eliminates external sources
2. Temperature 0.1 → eliminates variation
3. Max Tokens 500 → eliminates verbosity
4. GPT-4 Turbo → maximum obedience to the prompt📈 IMPACT ON THE MEDICAL SYSTEM:
✅ PRECISION: 100% deterministic answers ✅ SAFETY: Zero external information sources ✅ COST: Optimized for high volumes ✅ PERFORMANCE: Fast and concise answers ✅ SCALABILITY: Ready for intense usage
🎯 AI MODEL CONFIGURATION CONCLUSION:
OUR SETTINGS ENSURE:
"Medical Question" → "Search in Knowledge Base" → "Answer from the database" ↓ ↓ ↓ GPT-4 Turbo Qdrant + JSON Concise, precise Temperature 0.1 Active Sync Max 500 tokens
👉 This configuration transforms GPT-4 from a generalist model into a specialized medical search engine!
3. “EMBEDDINGS” TAB – DATABASE CONNECTION
📍 AI Engine → Chatbots → [Chatbot Name] → “Embeddings” Tab
🔗 ENVIRONMENT CONFIGURATION:
Environment: Quadrant (✅ Selected) Content Aware: No (❌ Disabled) - STRATEGIC DECISION
🎯 EMBEDDINGS CONFIGURATION ANALYSIS:
🏗️ ENVIRONMENT: QUADRANT – VECTOR ARCHITECTURE:
Selection: Quadrant (✅ Correct) Function: Connection to the cloud vector database Benefit: - Optimized storage for medical embeddings - Ultra-fast semantic search - Automatic scalability - Zero cost for the free plan (1GB)
🚫 CONTENT AWARE: NO – ISOLATION STRATEGY:
Status: ❌ No (Disabled) - CONSCIOUS DECISION WHAT IS CONTENT AWARE: - A function that adapts the chatbot to the content of the WordPress PAGE - Replaces {CONTENT} in the prompt with the text of the current page - Makes the chatbot "read" the page and respond in context WHY WE DISABLED IT: - Our medical chatbot MUST NOT know what is on the site's pages - It must function ISOLATEDLY, only with its medicine database - The {CONTENT} placeholder would introduce unwanted external information Outcome: We keep the chatbot STRICTLY FOCUSED on the medical database
🔍 CONTENT AWARE vs. COMPLETE ISOLATION COMPARISON:
❌ WITH CONTENT AWARE ENABLED:
User on page: "Acute cough treatment" Chatbot "reads" the page → understands the cough context Question: "what medicines do you recommend?" Answer: Is influenced by the page content PROBLEM: Mixes information from the page with the database
✅ WITH CONTENT AWARE DISABLED (OUR CONFIGURATION):
User on ANY page Chatbot COMPLETELY IGNORES the page content Question: "what medicines for cough?" Answer: Searches ONLY in the medicine database SUCCESS: Consistent answer regardless of the page where it is embedded
⚡ COMPLETE ISOLATION ARCHITECTURE:
🔄 DATA FLOW – TOTAL ISOLATION:
WORDPRESS PAGE: "Pancreatic Cancer Treatment" ↓ CHATBOT COMPLETELY IGNORES this content ↓ QUESTION: "what medicines for cancer?" ↓ SEARCHES DOAR în KNOWLEDGE BASE (medicamente JSON) ↓ RESPONSE: "I do not have relevant medicines for cancer"
🎯 BENEFITS OF ISOLATION:
✅ ABSOLUTE CONSISTENCY:
Same answer on any page of the site
Not influenced by page content
Predictable and verifiable behavior
✅ CLEAR SEPARATION:
Medical database = the only source of truth
Site content = only UI container
No mixing between the two
✅ INFORMATIONAL SECURITY:
Prevents “contamination” with information from pages
Ensures that answers come only from the verified source
Eliminates the risk of misinformation from the site content
🔧 STRATEGIC DECISION ANALYSIS:
🎯 WHY THIS SETUP IS CRITICAL:
🏥 PURE MEDICAL SPECIALIZATION:
The Chatbot is: "Medical Expert with its own Database" It is not: "Contextual Assistant that reads pages" Advantage: The same level of medical expertise everywhere
🔒 PROTECTION AGAINST CONTAMINATION:
If a page contained: incorrect medical information With Content Aware: the chatbot could be influenced Without Content Aware: the chatbot remains pure, only with its base
🌐 MAXIMUM PORTABILITY:
You can move the chatbot to any page
The behavior remains IDENTICAL
It does not need to be re-tested for every new page🎯 CONCLUSION – ISOLATED ARCHITECTURE:
🔬 OUR SYSTEM WORKS AS:
"Medical Expert in a Capsule" - Hermetically sealed from external influences - Powered only by its database - Pure and uncontaminated answers
👉 Disabled Content Aware ensures that the medical chatbot remains an independent and consistent entity, undistorted by site content!
