server.py

from fastapi import FastAPI
from fastapi.middleware.cors import CORSMiddleware
from fastapi.responses import StreamingResponse, FileResponse
from fastapi.staticfiles import StaticFiles
from pydantic import BaseModel
from langchain_community.llms import Ollama
from langchain.prompts import PromptTemplate
from typing import List, Dict, Optional, AsyncGenerator
import json
import asyncio
from pathlib import Path
import os
import re

# Import transformers for NER
try:
    from transformers import pipeline
    TRANSFORMERS_AVAILABLE = True
except ImportError:
    TRANSFORMERS_AVAILABLE = False
    print("Warning: transformers not available, NER models will not work")

app = FastAPI(title="Pub/Sub Multi-Agent System")

# Enable CORS
app.add_middleware(
    CORSMiddleware,
    allow_origins=["*"],
    allow_credentials=True,
    allow_methods=["*"],
    allow_headers=["*"],
)

# Mount static files
static_dir = Path(__file__).parent / "static"
static_dir.mkdir(exist_ok=True)
app.mount("/static", StaticFiles(directory=static_dir), name="static")

# Models
class DataSource(BaseModel):
    label: str
    content: str

class Agent(BaseModel):
    title: str
    prompt: str
    model: str
    subscribe_topic: str
    publish_topic: Optional[str] = None
    show_result: bool = False

class ExecutionRequest(BaseModel):
    data_sources: List[DataSource]
    user_question: str = ""
    agents: List[Agent]

# Pub/Sub Bus
class MessageBus:
    def __init__(self):
        self.subscribers: Dict[str, List[Agent]] = {}
        self.messages: Dict[str, str] = {}
        
    def reset(self):
        """Reset the bus for a new execution"""
        self.subscribers = {}
        self.messages = {}
    
    def _normalize_topic(self, topic: str) -> str:
        """Normalize topic to lowercase for case-insensitive matching"""
        return topic.lower().strip()
    
    def subscribe(self, topic: str, agent: Agent):
        """Subscribe an agent to a topic (case insensitive)"""
        normalized = self._normalize_topic(topic)
        if normalized not in self.subscribers:
            self.subscribers[normalized] = []
        self.subscribers[normalized].append(agent)
    
    def publish(self, topic: str, content: str):
        """Publish a message to a topic (case insensitive)"""
        normalized = self._normalize_topic(topic)
        self.messages[normalized] = content
    
    def get_message(self, topic: str) -> Optional[str]:
        """Get message from a topic (case insensitive)"""
        normalized = self._normalize_topic(topic)
        return self.messages.get(normalized)
    
    def get_subscribers(self, topic: str) -> List[Agent]:
        """Get all subscribers for a topic (case insensitive)"""
        normalized = self._normalize_topic(topic)
        return self.subscribers.get(normalized, [])

# Stream event helper
def create_event(event_type: str, **kwargs):
    data = {"type": event_type, **kwargs}
    return f"data: {json.dumps(data)}\n\n"

# Get LLM instance
def get_llm(model_name: str):
    return Ollama(model=model_name, temperature=0.1)

# NER pipeline cache
_ner_pipelines = {}

def get_ner_pipeline(model_name: str):
    """Get or create NER pipeline for the specified model"""
    if not TRANSFORMERS_AVAILABLE:
        raise RuntimeError("transformers package not available")
    
    if model_name not in _ner_pipelines:
        print(f"Loading NER model: {model_name}")
        _ner_pipelines[model_name] = pipeline(
            "ner",
            model=model_name,
            aggregation_strategy="simple"
        )
    return _ner_pipelines[model_name]

# Check if model is NER model
def is_ner_model(model_name: str) -> bool:
    """Check if the model is an NER model"""
    ner_models = [
        "samrawal/bert-base-uncased_clinical-ner",
        "OpenMed/OpenMed-NER-AnatomyDetect-BioPatient-108M"
    ]
    return model_name in ner_models

# Format NER output for display
def format_ner_result(text: str, entities: List[Dict]) -> str:
    """Format NER entities for human-readable display"""
    if not entities:
        return text
    
    # Sort entities by start position in reverse to avoid index issues
    sorted_entities = sorted(entities, key=lambda x: x['start'], reverse=True)
    
    result = text
    for entity in sorted_entities:
        start = entity['start']
        end = entity['end']
        entity_type = entity['entity_type']
        original_text = text[start:end]
        
        # Replace entity with labeled version
        labeled = f"[{original_text}:{entity_type}]"
        result = result[:start] + labeled + result[end:]
    
    return result

# Process NER with transformers pipeline
def process_ner(text: str, model_name: str) -> tuple[str, List[Dict]]:
    """Process text with NER pipeline and return JSON + formatted entities"""
    try:
        ner_pipeline = get_ner_pipeline(model_name)
        
        # Run NER
        entities = ner_pipeline(text)
        
        # Convert to our format with proper type conversion
        formatted_entities = []
        for entity in entities:
            formatted_entities.append({
                "text": str(entity['word']),
                "entity_type": str(entity['entity_group']),
                "start": int(entity['start']),
                "end": int(entity['end']),
                "score": float(entity.get('score', 0.0))  # Convert numpy float32 to Python float
            })
        
        # Create JSON output with proper serialization
        json_output = json.dumps(formatted_entities, indent=2)
        
        return json_output, formatted_entities
        
    except Exception as e:
        error_msg = f"NER processing failed: {str(e)}"
        return json.dumps({"error": error_msg}), []

# Execute agent
async def execute_agent(agent: Agent, input_content: str, data_sources: List[DataSource], user_question: str) -> tuple[str, Optional[List[Dict]], Optional[str]]:
    """Execute a single agent with the given input. Returns (result, entities, analyzed_text) where entities is for NER models."""
    
    # Case-insensitive replacement helper
    def replace_case_insensitive(text: str, placeholder: str, value: str) -> str:
        """Replace placeholder in text, case insensitive"""
        pattern = re.compile(re.escape(placeholder), re.IGNORECASE)
        return pattern.sub(value, text)
    
    # Start with the agent's prompt template
    prompt_text = agent.prompt if agent.prompt else ""
    
    # Replace standard placeholders (case insensitive)
    prompt_text = replace_case_insensitive(prompt_text, "{input}", input_content)
    prompt_text = replace_case_insensitive(prompt_text, "{question}", user_question)
    
    # Replace data source placeholders (case insensitive)
    for ds in data_sources:
        placeholder = "{" + ds.label + "}"
        prompt_text = replace_case_insensitive(prompt_text, placeholder, ds.content)
    
    # Check if this is an NER model
    if is_ner_model(agent.model):
        # For NER models, the rendered prompt IS the text to analyze
        text_to_analyze = prompt_text
        
        # If prompt is empty, use input content directly
        if not text_to_analyze.strip():
            text_to_analyze = input_content
        
        # Process with NER pipeline
        json_result, entities = process_ner(text_to_analyze, agent.model)
        
        # Return JSON result, entities, and the text that was analyzed
        return json_result, entities, text_to_analyze
    else:
        # Regular LLM processing
        llm = get_llm(agent.model)
        
        # Invoke LLM with the rendered prompt
        result = llm.invoke(prompt_text)
        
        return (result if isinstance(result, str) else str(result)), None, None

# Main execution pipeline
async def execute_pipeline(request: ExecutionRequest) -> AsyncGenerator[str, None]:
    try:
        bus = MessageBus()
        
        yield create_event("bus_init")
        
        # Reset and configure subscriptions
        bus.reset()
        
        # Subscribe all agents to their topics
        for agent in request.agents:
            if agent.subscribe_topic:
                bus.subscribe(agent.subscribe_topic, agent)
                yield create_event("agent_subscribed", agent=agent.title, topic=agent.subscribe_topic)
        
        # Publish START message
        start_message = request.user_question if request.user_question else "System initialized"
        bus.publish("START", start_message)
        yield create_event("message_published", topic="START", content=start_message)
        
        # Process messages in the bus
        processed_topics = set()
        max_iterations = 20  # Prevent infinite loops
        iteration = 0
        
        while iteration < max_iterations:
            iteration += 1
            
            # Find topics that have messages but haven't been processed
            topics_to_process = [topic for topic in bus.messages.keys() if topic not in processed_topics]
            
            if not topics_to_process:
                break
            
            for topic in topics_to_process:
                subscribers = bus.get_subscribers(topic)
                
                if not subscribers:
                    yield create_event("no_subscribers", topic=topic)
                    processed_topics.add(topic)
                    continue
                
                message_content = bus.get_message(topic)
                
                for agent in subscribers:
                    yield create_event("agent_triggered", agent=agent.title, topic=topic)
                    yield create_event("agent_processing", agent=agent.title)
                    yield create_event("agent_input", content=message_content)
                    
                    # Execute agent
                    try:
                        result, entities, analyzed_text = await execute_agent(agent, message_content, request.data_sources, request.user_question)
                        yield create_event("agent_output", content=result)
                        
                        # If agent wants to show result, send it to frontend
                        if agent.show_result:
                            yield create_event("show_result", agent=agent.title, content=result)
                            
                            # If this is an NER agent with entities, also send formatted NER result
                            if entities and is_ner_model(agent.model) and analyzed_text:
                                formatted_text = format_ner_result(analyzed_text, entities)
                                yield create_event("ner_result", agent=agent.title, formatted_text=formatted_text)
                        
                        # Publish result to agent's publish topic (if specified)
                        if agent.publish_topic:
                            bus.publish(agent.publish_topic, result)
                            yield create_event("message_published", topic=agent.publish_topic, content=result)
                        
                        yield create_event("agent_completed", agent=agent.title)
                        
                    except Exception as e:
                        yield create_event("error", message=f"Agent {agent.title} failed: {str(e)}")
                
                processed_topics.add(topic)
        
        yield create_event("execution_complete")
        
    except Exception as e:
        yield create_event("error", message=str(e))

@app.get("/")
async def root():
    """Serve the main web interface"""
    index_file = static_dir / "index.html"
    if index_file.exists():
        return FileResponse(index_file)
    return {"message": "Place index.html in the static/ directory"}

@app.post("/execute")
async def execute(request: ExecutionRequest):
    """Execute the pub/sub agent system with streaming logs"""
    return StreamingResponse(
        execute_pipeline(request),
        media_type="text/event-stream"
    )

@app.get("/health")
async def health():
    """Health check endpoint"""
    return {"status": "ok"}

if __name__ == "__main__":
    import uvicorn
    port = int(os.environ.get("PORT", 7860))
    uvicorn.run(app, host="0.0.0.0", port=port)