together_ai_llama_agent.py

import pandas as pd
import json
from typing import List, Literal, Optional
from pydantic import BaseModel
from dotenv import load_dotenv
from pydantic_ai import Agent
from csv_service import clean_data
from python_code_executor_service import PythonExecutor
from together_ai_instance_provider import InstanceProvider

load_dotenv()

instance_provider = InstanceProvider()

class CodeResponse(BaseModel):
    """Container for code-related responses"""
    language: str = "python"
    code: str

class ChartSpecification(BaseModel):
    """Details about requested charts"""
    image_description: str
    code: Optional[str] = None

class AnalysisOperation(BaseModel):
    """Container for a single analysis operation with its code and result"""
    code: CodeResponse
    result_var: str

class CsvChatResult(BaseModel):
    """Structured response for CSV-related AI interactions"""
    response_type: Literal["casual", "data_analysis", "visualization", "mixed"]
    
    # Casual chat response
    casual_response: str
    
    # Data analysis components
    analysis_operations: List[AnalysisOperation]
    
    # Visualization components
    charts: Optional[List[ChartSpecification]] = None
    
    
def get_csv_info(df: pd.DataFrame) -> dict:
    """Get metadata/info about the CSV"""
    info = {
        'num_rows': len(df),
        'num_cols': len(df.columns),
        'example_rows': df.head(2).to_dict('records'),
        'dtypes': {col: str(df[col].dtype) for col in df.columns},
        'columns': list(df.columns),
        'numeric_columns': [col for col in df.columns if pd.api.types.is_numeric_dtype(df[col])],
        'categorical_columns': [col for col in df.columns if pd.api.types.is_string_dtype(df[col])]
    }
    return info



def get_csv_system_prompt(df: pd.DataFrame) -> str:
    """Generate system prompt for CSV analysis"""
    csv_info = get_csv_info(df)
    
    prompt = f"""
You're a CSV analysis assistant. The pandas DataFrame is loaded as 'df' - use this variable.

CSV Info:
- Rows: {csv_info['num_rows']}, Cols: {csv_info['num_cols']}
- Columns: {csv_info['columns']}
- Sample: {csv_info['example_rows']}
- Dtypes: {csv_info['dtypes']}

Strict Rules:
1. Never recreate 'df' - use the existing variable
2. For analysis:
   - Complete code without imports
   - Use df directly (e.g., print(df[...].mean()))
3. For visualizations:
   - Create the most professional, publication-quality charts possible
   - Maximize descriptive elements and detail while maintaining clarity
   - Figure size: (14, 8) for complex charts, (12, 6) for simpler ones
   - Use comprehensive titles (fontsize=16) and axis labels (fontsize=14)
   - Include informative legends (fontsize=12) when appropriate
   - Add annotations for important data points where valuable
   - Rotate x-labels (45° if needed) with fontsize=12 for readability
   - Use colorblind-friendly palettes (seaborn 'deep', 'muted', or 'colorblind')
   - Add gridlines (alpha=0.3) when they improve readability
   - Include proper margins and padding to prevent label cutoff
   - For distributions, include kernel density estimates when appropriate
   - For time series, use appropriate date formatting and markers
   - Do not use any visualization library other than matplotlib or seaborn
   - Complete code with plt.tight_layout() before plt.show()
   - Example professional chart:
     plt.figure(figsize=(14, 8))
     ax = sns.barplot(x='category', y='value', data=df, palette='muted', ci=None)
     plt.title('Detailed Analysis of Values by Category', fontsize=16, pad=20)
     plt.xlabel('Category', fontsize=14)
     plt.ylabel('Average Value', fontsize=14)
     plt.xticks(rotation=45, ha='right', fontsize=12)
     plt.yticks(fontsize=12)
     ax.grid(True, linestyle='--', alpha=0.3)
     for p in ax.patches:
         ax.annotate(f'{{p.get_height():.1f}}', 
                    (p.get_x() + p.get_width() / 2., p.get_height()), 
                    ha='center', va='center', 
                    xytext=(0, 10), 
                    textcoords='offset points',
                    fontsize=12)
     plt.tight_layout()
     plt.show()
4. For Lists, Records, Tables, Dictionaries...etc for any data structure, always return them as JSON with correct indentation.

IMPORTANT: Code must be syntactically perfect and executable as-is.

"""
    return prompt


def create_csv_agent(df: pd.DataFrame, max_retries: int = 1) -> Agent:
    """Create and return a CSV analysis agent with API key rotation"""
    csv_system_prompt = get_csv_system_prompt(df)
    
    for attempt in range(max_retries):
        try:
            model = instance_provider.get_instance()
            if model is None:
                raise RuntimeError("No available API instances")
            
            csv_agent = Agent(
                model=model,
                output_type=CsvChatResult,
                system_prompt=csv_system_prompt,
            )
        
            return csv_agent
            
        except Exception as e:
            api_key = instance_provider.get_api_key_for_model(model)
            if api_key:
                print(f"Error with API key (attempt {attempt + 1}): {str(e)}")
                instance_provider.report_error(api_key)
            continue
    
    raise RuntimeError(f"Failed to create agent after {max_retries} attempts")


async def query_csv_agent(csv_url: str, question: str, chat_id: str) -> str:
    """Query the CSV agent with a DataFrame and question and return formatted output"""
    
    # Get the DataFrame from the CSV URL
    df = clean_data(csv_url)
    
    # Create agent and get response
    agent = create_csv_agent(df)
    result = await agent.run(question)
    
    # Process the response through PythonExecutor
    executor = PythonExecutor(df)
    
    # Convert the raw output to CsvChatResult if needed
    if not isinstance(result.output, CsvChatResult):
        # Handle case where output needs conversion
        try:
            response_data = result.output if isinstance(result.output, dict) else json.loads(result.output)
            chat_result = CsvChatResult(**response_data)
        except Exception as e:
            raise ValueError(f"Could not parse agent response: {str(e)}")
    else:
        chat_result = result.output
        
        print("Chat Result Original Object:", chat_result)
    
    # Process and format the response
    formatted_output = await executor.process_response(chat_result, chat_id)
    
    return formatted_output