# Domain-Specific System Prompts in meeTARA

## Current Architecture

### 1. **Model Level (Core) - System Prompt**

**Location:** `core/meetara_lab_core.py` → `_get_system_prompt()`
**Config:** `config/agent_config.json` → `model.system_prompt.prompt`

**Current Behavior:**
- **ONE global system prompt** used for ALL queries
- Defined in `agent_config.json`:
```json
{
  "model": {
    "system_prompt": {
      "enabled": true,
      "prompt": "You are meeTARA, an accurate AI assistant.\nIMPORTANT: For math/calculations, always verify your answer before responding.\nRespond in this format:\n🎯 **Answer**: Direct answer (must be correct)\n📊 **Details**: Key facts (2-3 bullets)\n⚡ **Steps**: Show your work/calculation\n💡 **Note**: Any warnings (1 sentence)\nBe accurate first, concise second."
    }
  }
}
```

- Applied in ChatML format:
```
<|im_start|>system
{system_prompt}
<|im_end|>
<|im_start|>user
{query}
<|im_end|>
<|im_start|>assistant
```

**Limitation:** Same system prompt for all domains (healthcare, business, education, etc.)

---

### 2. **Agent Level - Enhanced Instructions**

**Location:** `core/meetara_agent.py` → `run()` method (line ~1384)
**Config:** `config/agent_config.json` → `web_search.news_query_instructions`

**Current Behavior:**
- **Tool-specific instructions** added to user prompt (NOT system prompt)
- Example: News query instructions added when web search is used for news:
```json
{
  "web_search": {
    "news_query_instructions": {
      "enabled": true,
      "instruction_template": "CRITICAL INSTRUCTIONS FOR NEWS/HEADLINES QUERY:\n1. Extract and list the ACTUAL NEWS HEADLINES..."
    }
  }
}
```

- Applied as part of enhanced user prompt:
```
{query}

=== Tool Results ===
{search_results}

{instruction_text}  ← Added here (in user message, not system)
```

**Current Domain Support:**
- ✅ News/Headlines queries (via `news_query_instructions`)
- ❌ No other domain-specific instructions currently

---

## Domain Mapping Structure

**Location:** `config/domain_mapping.json`

**Available Domain Categories:**
1. `healthcare` (🏥) - Medical, wellness, health
2. `daily_life` (🏠) - Personal, family, everyday
3. `business` (💼) - Business, entrepreneurship
4. `education` (📚) - Learning, skill development
5. `creative` (🎨) - Creative expression, arts
6. `psychology_wellness` (🧘) - Mental wellness
7. `sports_recreation` (⚽) - Physical fitness
8. `technology` (💻) - Technology, digital solutions
9. `legal_financial` (⚖️) - Legal matters, finance
10. `travel_tourism` (🧳) - Travel planning
11. `food_cooking` (🍳) - Culinary arts
12. And more...

**Domain Properties (Available but NOT Currently Used for System Prompts):**
- `keywords` - Keywords to detect domain
- `emotional_context` - Tone/style (e.g., "compassionate", "professional")
- `description` - Domain description
- `model_preference` - Model type preference

---

## Proposed: Domain-Specific System Prompts

### Option 1: Agent-Level Domain Detection + Enhanced Instructions

**How it works:**
1. Agent detects domain from query keywords (using `domain_mapping.json`)
2. Agent adds domain-specific instructions to user prompt (similar to news queries)
3. System prompt remains global

**Pros:**
- ✅ Easy to implement
- ✅ Non-invasive (doesn't change core)
- ✅ Similar to existing news query pattern

**Cons:**
- ❌ Instructions in user prompt (weaker than system prompt)
- ❌ Not as effective as system-level prompts

### Option 2: Core-Level Domain-Specific System Prompts

**How it works:**
1. Agent detects domain and passes domain info to Core
2. Core selects domain-specific system prompt from config
3. Core applies domain-specific system prompt to model

**Pros:**
- ✅ More effective (system prompt is stronger than user instructions)
- ✅ Proper separation (domain detection in agent, prompt selection in core)
- ✅ Domain-specific behavior baked into model context

**Cons:**
- ❌ Requires changes to Core API
- ❌ More complex implementation

---

## Implementation Recommendations

### Recommended Approach: **Hybrid (Option 1 + Option 2)**

1. **Agent Level:**
   - Detect domain from query keywords (using `domain_mapping.json`)
   - Add domain-specific context instructions to user prompt (for tool results)

2. **Core Level:**
   - Add domain-specific system prompt enhancement
   - Optionally prepend domain context to base system prompt

**Example Domain System Prompts Structure:**
```json
{
  "model": {
    "system_prompt": {
      "enabled": true,
      "base_prompt": "You are meeTARA, an accurate AI assistant...",
      "domain_specific_prompts": {
        "healthcare": {
          "enabled": true,
          "enhancement": "You are responding to a healthcare/medical query. Be compassionate, professional, and include appropriate medical disclaimers. Always remind users to consult healthcare professionals for medical advice.",
          "append_mode": true
        },
        "business": {
          "enabled": true,
          "enhancement": "You are responding to a business query. Be professional, data-driven, and focus on actionable insights. Include relevant business metrics and considerations.",
          "append_mode": true
        },
        "legal_financial": {
          "enabled": true,
          "enhancement": "You are responding to a legal or financial query. Be precise, cite sources when possible, and include appropriate disclaimers about not providing legal/financial advice.",
          "append_mode": true
        }
      }
    }
  }
}
```

**Flow:**
```
User Query → Agent detects domain → 
  → Agent: Adds domain-specific instructions to user prompt (for tool results)
  → Core: Enhances system prompt with domain context
  → Model: Generates response with domain-aware context
```

---

## Next Steps

To implement domain-specific system prompts:

1. **Add domain detection function** (uses `domain_mapping.json` keywords)
2. **Extend `agent_config.json`** with `domain_specific_prompts` section
3. **Modify Core `_get_system_prompt()`** to accept domain parameter
4. **Modify Agent `run()`** to:
   - Detect domain
   - Pass domain to Core
   - Add domain-specific instructions to user prompt (for tool results)

Would you like me to implement this?