backend/api/routes/agent.py

# =============================================================
# File: backend/api/routes/agent.py
# =============================================================

from fastapi import APIRouter
from fastapi.responses import StreamingResponse
from pydantic import BaseModel
import os
import sys
import json
from pathlib import Path
from typing import AsyncGenerator

# Add backend to path for imports
backend_dir = Path(__file__).parent.parent.parent
sys.path.insert(0, str(backend_dir))

from api.services.agent_orchestrator import AgentOrchestrator
from api.models.agent import AgentRequest, AgentResponse


router = APIRouter()


orchestrator = AgentOrchestrator(
    rag_mcp_url=os.getenv("RAG_MCP_URL", "http://localhost:8001"),
    web_mcp_url=os.getenv("WEB_MCP_URL", "http://localhost:8002"),
    admin_mcp_url=os.getenv("ADMIN_MCP_URL", "http://localhost:8003"),
    llm_backend=os.getenv("LLM_BACKEND", "ollama")
)


class ChatRequest(BaseModel):
    tenant_id: str
    user_id: str | None = None
    message: str
    conversation_history: list[dict] = []
    temperature: float = 0.0


@router.post("/message", response_model=AgentResponse)
async def agent_chat(req: ChatRequest):
    agent_req = AgentRequest(
        tenant_id=req.tenant_id,
        user_id=req.user_id,
        message=req.message,
        conversation_history=req.conversation_history,
        temperature=req.temperature
    )
    return await orchestrator.handle(agent_req)


@router.post("/message/stream")
async def agent_chat_stream(req: ChatRequest):
    """Stream agent response word by word using Server-Sent Events."""
    agent_req = AgentRequest(
        tenant_id=req.tenant_id,
        user_id=req.user_id,
        message=req.message,
        conversation_history=req.conversation_history,
        temperature=req.temperature
    )
    
    async def generate_stream() -> AsyncGenerator[str, None]:
        """Generate streaming response."""
        try:
            # FIRST: Check admin rules - if any rule matches, respond according to rule
            yield f"data: {json.dumps({'status': 'processing', 'message': 'Checking rules...'})}\n\n"
            matches = await orchestrator.redflag.check(agent_req.tenant_id, agent_req.message)
            
            if matches:
                # Categorize rules: brief response rules vs blocking rules
                brief_response_rules = []
                blocking_rules = []
                
                for match in matches:
                    rule_text = (match.description or match.pattern or "").lower()
                    is_brief_rule = (
                        match.severity == "low" and (
                            "greeting" in rule_text or 
                            "brief" in rule_text or 
                            "simple response" in rule_text or
                            "keep.*response.*brief" in rule_text or
                            "do not.*verbose" in rule_text or
                            "respond.*briefly" in rule_text
                        )
                    )
                    
                    if is_brief_rule:
                        brief_response_rules.append(match)
                    else:
                        blocking_rules.append(match)
                
                # Handle brief response rules (greetings, etc.) - return immediately
                if brief_response_rules and not blocking_rules:
                    brief_responses = [
                        "Hello! How can I help you today?",
                        "Hi there! What can I assist you with?",
                        "Hello! I'm here to help. What do you need?",
                        "Hi! How can I assist you?"
                    ]
                    import random
                    brief_response = random.choice(brief_responses)
                    
                    # Stream the brief response word by word
                    yield f"data: {json.dumps({'status': 'streaming', 'message': ''})}\n\n"
                    words = brief_response.split()
                    for word in words:
                        yield f"data: {json.dumps({'token': word + ' ', 'done': False})}\n\n"
                    yield f"data: {json.dumps({'token': '', 'done': True})}\n\n"
                    return
                
                # Handle blocking rules (security, compliance, etc.)
                if blocking_rules:
                    matches = blocking_rules
            
            if matches:
                # For red flags, generate streaming response via LLM
                violations_details = []
                for i, m in enumerate(matches, 1):
                    rule_name = m.description or m.pattern or "Policy violation"
                    detail = f"{i}. **{rule_name}** (Severity: {m.severity})"
                    if m.matched_text:
                        detail += f"\n   - Detected phrase: \"{m.matched_text}\""
                    violations_details.append(detail)
                
                llm_prompt = f"""A user made the following request: "{agent_req.message}"

However, this request violates company policies. The following policy violations were detected:

{chr(10).join(violations_details)}

Your task: Write a clear, professional, and empathetic response to inform the user that:
1. Their request cannot be processed due to policy violations
2. Which specific policy was violated (mention it naturally)
3. The incident has been logged for security review
4. They should contact an administrator if they need assistance or believe this is an error

Write a natural, conversational response (2-4 sentences) that feels helpful rather than robotic. Be professional but understanding.

Response:"""
                
                async for token in orchestrator.llm.stream_call(llm_prompt, agent_req.temperature):
                    yield f"data: {json.dumps({'token': token, 'done': False})}\n\n"
                
                yield f"data: {json.dumps({'token': '', 'done': True})}\n\n"
                return
            
            # STEP 2: ONLY IF NO RULES MATCHED - Proceed with normal flow
            yield f"data: {json.dumps({'status': 'classifying', 'message': 'Understanding your question...'})}\n\n"
            
            # Check if this is an admin identity question - handle it specially
            user_text = agent_req.message.lower().strip()
            user_text_normalized = " ".join(user_text.split())
            admin_phrases = [
                "who is the admin",
                "who's the admin",
                "who is admin",
                "who is the administrator",
                "who administers this platform",
                "who is the owner",
                "who owns this platform",
                "who is the admin of integrachat",
                "who administers integrachat",
            ]
            is_admin_question = (
                any(p in user_text_normalized for p in admin_phrases) or
                ("who" in user_text and "admin" in user_text)
            )
            
            # For admin questions, ALWAYS check RAG first and answer directly from knowledge base
            if is_admin_question:
                yield f"data: {json.dumps({'status': 'searching', 'message': 'Searching knowledge base for admin information...'})}\n\n"
                try:
                    rag_prefetch = await orchestrator.mcp.call_rag(agent_req.tenant_id, agent_req.message)
                    rag_results = []
                    if isinstance(rag_prefetch, dict):
                        rag_results = rag_prefetch.get("results") or rag_prefetch.get("hits") or []
                    
                    # If we have RAG hits, return the answer directly from the knowledge base
                    if rag_results:
                        best_hit = rag_results[0]
                        admin_text = best_hit.get("text") or best_hit.get("content") or str(best_hit)
                        response_text = f"According to the tenant knowledge base, {admin_text.strip()}"
                    else:
                        response_text = "I don't know who administers this platform based on the tenant data."
                    
                    # Stream the response word by word
                    yield f"data: {json.dumps({'status': 'streaming', 'message': ''})}\n\n"
                    import asyncio
                    words = response_text.split()
                    for word in words:
                        yield f"data: {json.dumps({'token': word + ' ', 'done': False})}\n\n"
                        await asyncio.sleep(0)
                    yield f"data: {json.dumps({'token': '', 'done': True})}\n\n"
                    return
                except Exception as rag_err:
                    # If RAG fails, fall through to normal flow
                    pass
            
            intent = await orchestrator.intent.classify(agent_req.message)
            
            # Pre-fetch RAG if needed (for non-admin questions)
            rag_results = []
            if intent == "rag" or "rag" in intent.lower():
                yield f"data: {json.dumps({'status': 'searching', 'message': 'Searching knowledge base...'})}\n\n"
                try:
                    rag_prefetch = await orchestrator.mcp.call_rag(agent_req.tenant_id, agent_req.message)
                    if isinstance(rag_prefetch, dict):
                        rag_results = rag_prefetch.get("results") or rag_prefetch.get("hits") or []
                except Exception:
                    pass
            
            # Also check if we have prefetched RAG results from earlier (for all questions)
            # This ensures RAG context is used even if intent isn't "rag"
            if not rag_results:
                try:
                    rag_prefetch = await orchestrator.mcp.call_rag(agent_req.tenant_id, agent_req.message)
                    if isinstance(rag_prefetch, dict):
                        rag_results = rag_prefetch.get("results") or rag_prefetch.get("hits") or []
                except Exception:
                    pass
            
            # Build prompt with context
            if rag_results:
                context = "\n\n".join([r.get("text", "")[:500] for r in rag_results[:3]])
                prompt = f"""Based on the following context, answer the user's question:

Context:
{context}

User's question: {agent_req.message}

Answer:"""
            else:
                prompt = agent_req.message
            
            # Signal that streaming is starting
            yield f"data: {json.dumps({'status': 'streaming', 'message': ''})}\n\n"
            
            # Stream LLM response - flush each token immediately
            # Import asyncio for potential delays if needed
            import asyncio
            async for token in orchestrator.llm.stream_call(prompt, agent_req.temperature):
                if token:  # Only send non-empty tokens
                    yield f"data: {json.dumps({'token': token, 'done': False})}\n\n"
                    # Small delay to ensure proper flushing (optional, can remove if not needed)
                    await asyncio.sleep(0)  # Yield control to event loop
            
            yield f"data: {json.dumps({'token': '', 'done': True})}\n\n"
            
        except Exception as e:
            error_msg = json.dumps({'error': str(e), 'done': True})
            yield f"data: {error_msg}\n\n"
    
    return StreamingResponse(
        generate_stream(),
        media_type="text/event-stream",
        headers={
            "Cache-Control": "no-cache",
            "Connection": "keep-alive",
            "X-Accel-Buffering": "no"
        }
    )


@router.post("/debug")
async def agent_debug(req: ChatRequest):
    """
    Returns detailed debugging information about agent reasoning.
    Includes intent classification, tool selection, reasoning trace, and tool traces.
    """
    agent_req = AgentRequest(
        tenant_id=req.tenant_id,
        user_id=req.user_id,
        message=req.message,
        conversation_history=req.conversation_history,
        temperature=req.temperature
    )
    response = await orchestrator.handle(agent_req)
    
    return {
        "request": {
            "tenant_id": req.tenant_id,
            "user_id": req.user_id,
            "message": req.message[:200],
            "temperature": req.temperature
        },
        "response": {
            "text": response.text[:500] + "..." if len(response.text) > 500 else response.text,
            "decision": response.decision.dict() if response.decision else None,
            "tool_traces": response.tool_traces,
            "reasoning_trace": response.reasoning_trace
        },
        "debug_info": {
            "intent": response.reasoning_trace[1].get("intent") if len(response.reasoning_trace) > 1 else None,
            "tool_selection": next((t for t in response.reasoning_trace if t.get("step") == "tool_selection"), None),
            "tool_scores": next((t for t in response.reasoning_trace if t.get("step") == "tool_scoring"), None),
            "redflag_check": next((t for t in response.reasoning_trace if t.get("step") == "redflag_check"), None),
            "total_steps": len(response.reasoning_trace)
        }
    }


@router.post("/plan")
async def agent_plan(req: ChatRequest):
    """
    Returns only the agent's planning output (tool selection decision).
    Useful for understanding what tools the agent would use without executing them.
    """
    from ..services.intent_classifier import IntentClassifier
    from ..services.tool_selector import ToolSelector
    from ..services.tool_scoring import ToolScoringService
    import os
    
    # Create minimal orchestrator components for planning only
    llm = orchestrator.llm
    intent_classifier = IntentClassifier(llm_client=llm)
    tool_selector = ToolSelector(llm_client=llm)
    tool_scorer = ToolScoringService()
    
    # Classify intent
    intent = await intent_classifier.classify(req.message)
    
    # Pre-fetch RAG for context (optional)
    rag_results = []
    try:
        rag_prefetch = await orchestrator.mcp.call_rag(req.tenant_id, req.message)
        if isinstance(rag_prefetch, dict):
            rag_results = rag_prefetch.get("results") or rag_prefetch.get("hits") or []
    except Exception:
        pass
    
    # Score tools
    tool_scores = tool_scorer.score(req.message, intent, rag_results)
    
    # Select tools
    ctx = {
        "tenant_id": req.tenant_id,
        "rag_results": rag_results,
        "tool_scores": tool_scores
    }
    decision = await tool_selector.select(intent, req.message, ctx)
    
    return {
        "tenant_id": req.tenant_id,
        "message": req.message,
        "intent": intent,
        "tool_scores": tool_scores,
        "plan": decision.dict(),
        "steps": decision.tool_input.get("steps", []) if decision.tool_input else [],
        "reason": decision.reason
    }