# Comment Processing with Agentic Workflow

A scalable, modular system for processing comments from multiple data sources using OpenAI API, LangChain, and LangGraph. The system performs language detection, translation, and context-aware sentiment analysis using an agentic workflow architecture.

## Data Sources Supported

- **Social Media Comments**: External platforms (Facebook, Instagram, YouTube, etc.)
- **Musora Internal Comments**: Comments from Musora internal applications
- **Extensible Architecture**: Easily add new data sources via configuration

## Features

- **Multi-Source Support**: Process comments from multiple data sources with a single codebase
- **Configuration-Driven**: Add new data sources without code changes
- **Parent Comment Context**: Automatically includes parent comment text for reply analysis
- **Modular Agent Architecture**: Extensible base classes for easy addition of new agents
- **Language Detection**: Hybrid approach using lingua library for fast English detection, with LLM fallback for non-English languages
- **Translation**: High-quality translation for non-English comments using OpenAI models
- **Context-Aware Sentiment Analysis**:
  - Uses content description for context
  - Includes parent comment text when analyzing replies
  - Multi-label intent classification
- **LangGraph Workflow**: Flexible graph-based orchestration of agent operations
- **Snowflake Integration**: Seamless data fetching and storage with source-specific tables
- **Parallel Processing**: Multiprocessing support for high-performance batch processing
- **Dynamic Batch Sizing**: Intelligent batch size calculation based on workload and available resources
- **Independent Batch Execution**: Each batch processes and stores results independently
- **Comprehensive Logging**: Detailed logging for monitoring and debugging
- **Scalable Configuration**: Easy-to-modify sentiment categories and intents via JSON config

## Project Structure

```
musora-sentiment-analysis/
├── agents/
│   ├── __init__.py
│   ├── base_agent.py                    # Base class for all agents
│   ├── language_detection_agent.py      # Language detection agent
│   ├── translation_agent.py             # Translation agent
│   └── sentiment_analysis_agent.py      # Sentiment analysis agent (parent context support)
├── workflow/
│   ├── __init__.py
│   └── comment_processor.py             # LangGraph workflow orchestrator
├── sql/
│   ├── fetch_comments.sql               # Query for social media comments (with parent join)
│   ├── fetch_musora_comments.sql        # Query for Musora internal comments (with parent join)
│   ├── create_ml_features_table.sql     # Schema for social media table (with parent fields)
│   ├── init_musora_table.sql            # Initialize empty Musora table (run first!)
│   └── create_musora_ml_features_table.sql  # Full Musora schema with views (optional)
├── config_files/
│   ├── data_sources_config.json         # Data source configuration (NEW)
│   ├── sentiment_config.json            # Configuration for agents and workflow
│   └── sentiment_analysis_config.json   # Sentiment categories and intents
├── logs/                                 # Processing logs (auto-created)
├── LLM.py                               # LLM utility class
├── SnowFlakeConnection.py               # Snowflake connection handler
├── main.py                              # Main execution script (multi-source support)
├── requirements.txt                      # Python dependencies
├── .env                                 # Environment variables (not in git)
├── README.md                            # This file
└── CLAUDE.md                            # Detailed technical documentation
```

## Setup

### 1. Install Dependencies

```bash
pip install -r requirements.txt
```

### 2. Configure Environment Variables

Ensure your `.env` file contains the required credentials:

```env
# Snowflake
SNOWFLAKE_USER=your_user
SNOWFLAKE_PASSWORD=your_password
SNOWFLAKE_ACCOUNT=your_account
SNOWFLAKE_ROLE=your_role
SNOWFLAKE_DATABASE=SOCIAL_MEDIA_DB
SNOWFLAKE_WAREHOUSE=your_warehouse
SNOWFLAKE_SCHEMA=ML_FEATURES

# OpenAI
OPENAI_API_KEY=your_openai_key
```

### 3. Create Snowflake Tables

Run the SQL scripts to create the output tables:

```bash
# Execute the SQL files in Snowflake
# For social media comments (if not already exists)
sql/create_ml_features_table.sql

# For Musora internal comments - INITIAL SETUP (First time only)
# This creates the empty table structure
sql/init_musora_table.sql
```

**Note**: Run `init_musora_table.sql` before the first Musora comments processing run. After that, you can optionally run `create_musora_ml_features_table.sql` to create the additional views if needed.

## Usage

### Basic Usage (Process All Data Sources)

Process unprocessed comments from all enabled data sources:

```bash
python main.py
```

This will:
- Process all enabled data sources (social media and Musora comments)
- Fetch only comments that haven't been processed yet
- Process them through the workflow using parallel workers (CPU count - 2, max 5)
- Each batch processes and stores to Snowflake independently
- Append new results to the existing tables (no overwrite)

### Process Specific Data Source

Process only social media comments:

```bash
python main.py --data-source social_media
```

Process only Musora internal comments:

```bash
python main.py --data-source musora_comments
```

### Process Limited Number of Comments

Limit applies per data source:

```bash
# Process 100 comments from each enabled data source
python main.py --limit 100

# Process 100 comments from only Musora source
python main.py --limit 100 --data-source musora_comments
```

### Sequential Processing (Debug Mode)

For debugging purposes, use sequential processing:

```bash
python main.py --limit 100 --sequential
```

This processes all comments in a single batch, making it easier to debug issues.

### First Run for New Data Source

For the first run of Musora comments:

1. **First**: Run the initialization SQL script in Snowflake:
   ```sql
   -- Execute in Snowflake
   sql/init_musora_table.sql
   ```

2. **Then**: Run the processing with overwrite flag:
   ```bash
   python main.py --overwrite --data-source musora_comments --limit 100
   ```

**Why two steps?**
- The fetch query checks for already-processed comments by querying the output table
- On first run, that table doesn't exist, causing an error
- The init script creates the empty table structure first
- Then processing can run normally

**Warning**: Overwrite will replace all existing data in the output table. Only use for initial table creation or when reprocessing from scratch.

### Custom Configuration File

```bash
python main.py --config path/to/custom_config.json
```

### Command-Line Arguments

- `--limit N`: Process only N comments per data source (default: 10000)
- `--overwrite`: Overwrite existing Snowflake table (default: append mode)
- `--config PATH`: Custom configuration file path
- `--sequential`: Use sequential processing instead of parallel (for debugging)
- `--data-source SOURCE`: Process only specific data source (e.g., social_media, musora_comments)

### Parallel Processing

The system uses multiprocessing to process comments in parallel:

**Worker Calculation**:
- Number of workers: `CPU count - 2` (max 5 workers)
- Leaves CPU cores available for system operations
- Example: 8-core system → 5 workers (capped at max)

**Dynamic Batch Sizing**:
- Batch size calculated as: `total_comments / num_workers`
- Minimum batch size: 20 comments
- Maximum batch size: 1000 comments
- Batches ≤ 20 comments are not split

**Independent Execution**:
- Each batch runs in a separate process
- Batches store to Snowflake immediately upon completion
- No waiting for all batches to complete
- Failed batches don't affect successful ones

**Performance**:
- Expected speedup: ~1.8-4.5x depending on number of workers
- Real-time progress reporting as batches complete
- Processing time and average per comment displayed in summary

### Incremental Processing

The pipeline is designed for incremental processing:
- **Automatic deduplication**: SQL query excludes comments already in `COMMENT_SENTIMENT_FEATURES`
- **Append-only by default**: New results are added without overwriting existing data
- **Failed comment retry**: Comments with `success=False` are not stored and will be retried in future runs
- **Run regularly**: Safe to run daily/weekly to process new comments

## Configuration

### Data Sources Configuration

The `config_files/data_sources_config.json` file defines available data sources:

```json
{
    "data_sources": {
        "social_media": {
            "name": "Social Media Comments",
            "enabled": true,
            "sql_query_file": "sql/fetch_comments.sql",
            "output_config": {
                "table_name": "COMMENT_SENTIMENT_FEATURES",
                "database": "SOCIAL_MEDIA_DB",
                "schema": "ML_FEATURES"
            }
        },
        "musora_comments": {
            "name": "Musora Internal Comments",
            "enabled": true,
            "sql_query_file": "sql/fetch_musora_comments.sql",
            "output_config": {
                "table_name": "MUSORA_COMMENT_SENTIMENT_FEATURES",
                "database": "SOCIAL_MEDIA_DB",
                "schema": "ML_FEATURES"
            },
            "additional_fields": [
                "PERMALINK_URL",
                "THUMBNAIL_URL"
            ]
        }
    }
}
```

**To add a new data source**: Simply add a new entry to this config file and create the corresponding SQL query file.

### Agent Configuration

The `config_files/sentiment_config.json` file controls agent behavior:

```json
{
    "agents": {
        "language_detection": {
            "model": "gpt-5-nano",
            "temperature": 0.0,
            "max_retries": 3
        },
        "translation": {
            "model": "gpt-5-nano",
            "temperature": 0.3,
            "max_retries": 3
        },
        "sentiment_analysis": {
            "model": "gpt-5-nano",
            "temperature": 0.2,
            "max_retries": 3
        }
    },
    "workflow": {
        "description": "Batch size is calculated dynamically based on number of workers (min: 20, max: 1000)",
        "parallel_processing": {
            "enabled": true,
            "worker_calculation": "CPU count - 2, max 5 workers",
            "min_batch_size": 20,
            "max_batch_size": 1000
        }
    },
    "snowflake": {
        "output_table": "COMMENT_SENTIMENT_FEATURES",
        "database": "SOCIAL_MEDIA_DB",
        "schema": "ML_FEATURES"
    }
}
```

**Note**: Batch size is now calculated dynamically and no longer needs to be configured manually.

### Sentiment Categories Configuration

The `config_files/sentiment_analysis_config.json` file defines sentiment categories and intents (easily extensible):

```json
{
    "sentiment_polarity": {
        "categories": [
            {"value": "very_positive", "label": "Very Positive", "description": "..."},
            {"value": "positive", "label": "Positive", "description": "..."},
            {"value": "neutral", "label": "Neutral", "description": "..."},
            {"value": "negative", "label": "Negative", "description": "..."},
            {"value": "very_negative", "label": "Very Negative", "description": "..."}
        ]
    },
    "intent": {
        "categories": [
            {"value": "praise", "label": "Praise", "description": "..."},
            {"value": "question", "label": "Question", "description": "..."},
            {"value": "request", "label": "Request", "description": "..."},
            {"value": "feedback_negative", "label": "Negative Feedback", "description": "..."},
            {"value": "suggestion", "label": "Suggestion", "description": "..."},
            {"value": "humor_sarcasm", "label": "Humor/Sarcasm", "description": "..."},
            {"value": "off_topic", "label": "Off Topic", "description": "..."},
            {"value": "spam_selfpromo", "label": "Spam/Self-Promotion", "description": "..."}
        ]
    },
    "reply_policy": {
        "requires_reply_intents": ["question", "request"],
        "description": "Comments with these intents should be flagged for reply"
    },
    "intent_settings": {
        "multi_label": true,
        "description": "Intent can have multiple labels as a comment can express multiple intents"
    }
}
```

## Adding New Agents

The system is designed for easy extensibility. To add a new agent:

### 1. Create Agent Class

```python
from agents.base_agent import BaseAgent
from typing import Dict, Any

class MyNewAgent(BaseAgent):
    def __init__(self, config: Dict[str, Any], api_key: str):
        super().__init__("MyNewAgent", config)
        # Initialize your agent-specific components

    def validate_input(self, input_data: Dict[str, Any]) -> bool:
        # Validate input data
        return True

    def process(self, input_data: Dict[str, Any]) -> Dict[str, Any]:
        # Implement your agent logic
        return {"success": True, "result": "..."}
```

### 2. Update Workflow

Add the agent to `workflow/comment_processor.py`:

```python
# Add to CommentState TypedDict
new_agent_result: str

# Add node
workflow.add_node("my_new_agent", self._my_new_agent_node)

# Add edges
workflow.add_edge("translation", "my_new_agent")
workflow.add_edge("my_new_agent", END)
```

### 3. Update Configuration

Add agent config to `sentiment_config.json`:

```json
{
    "agents": {
        "my_new_agent": {
            "name": "MyNewAgent",
            "model": "gpt-4o-mini",
            "temperature": 0.5,
            "max_retries": 3
        }
    }
}
```

## Output Schema

### Social Media Comments Table
Stored in `SOCIAL_MEDIA_DB.ML_FEATURES.COMMENT_SENTIMENT_FEATURES`

### Musora Comments Table
Stored in `SOCIAL_MEDIA_DB.ML_FEATURES.MUSORA_COMMENT_SENTIMENT_FEATURES`

### Common Columns (Both Tables)

| Column | Type | Description |
|--------|------|-------------|
| COMMENT_SK | NUMBER | Surrogate key from source |
| COMMENT_ID | VARCHAR | Platform comment ID |
| ORIGINAL_TEXT | VARCHAR | Original comment text |
| **PARENT_COMMENT_ID** | **VARCHAR** | **ID of parent comment if this is a reply** |
| **PARENT_COMMENT_TEXT** | **VARCHAR** | **Text of parent comment for context** |
| DETECTED_LANGUAGE | VARCHAR | Detected language name |
| LANGUAGE_CODE | VARCHAR | ISO 639-1 code |
| IS_ENGLISH | BOOLEAN | Is comment in English |
| TRANSLATED_TEXT | VARCHAR | English translation |
| TRANSLATION_PERFORMED | BOOLEAN | Was translation performed |
| SENTIMENT_POLARITY | VARCHAR | Sentiment (very_positive, positive, neutral, negative, very_negative) |
| INTENT | VARCHAR | Multi-label intents (comma-separated) |
| REQUIRES_REPLY | BOOLEAN | Does comment need a response |
| SENTIMENT_CONFIDENCE | VARCHAR | Analysis confidence (high, medium, low) |
| PROCESSING_SUCCESS | BOOLEAN | Processing status |
| PROCESSED_AT | TIMESTAMP | Processing timestamp |

### Musora-Specific Additional Columns

| Column | Type | Description |
|--------|------|-------------|
| PERMALINK_URL | VARCHAR | Web URL path of the content |
| THUMBNAIL_URL | VARCHAR | Thumbnail URL of the content |

### Available Views

**Social Media:**
- `VW_COMMENTS_REQUIRING_REPLY`: Comments that need responses (includes parent comment info)
- `VW_SENTIMENT_DISTRIBUTION`: Sentiment and intent statistics by channel (includes reply comment count)
- `VW_NON_ENGLISH_COMMENTS`: Filtered view of non-English comments

**Musora:**
- `VW_MUSORA_COMMENTS_REQUIRING_REPLY`: Musora comments needing responses
- `VW_MUSORA_SENTIMENT_DISTRIBUTION`: Musora sentiment and intent statistics
- `VW_MUSORA_NON_ENGLISH_COMMENTS`: Non-English Musora comments

## Workflow Architecture

The system uses LangGraph to create a flexible, state-based workflow:

```
┌─────────────────────┐
│  Fetch Comments     │
│  from Snowflake     │
│  (Unprocessed Only) │
└──────────┬──────────┘
           │
           ▼
┌─────────────────────┐
│  Language Detection │
│  Agent              │
└──────────┬──────────┘
           │
           ▼
      ┌────┴────┐
      │ English?│
      └────┬────┘
           │
     ┌─────┴─────┐
     │           │
   Yes          No
     │           │
     │           ▼
     │    ┌─────────────┐
     │    │ Translation │
     │    │ Agent       │
     │    └──────┬──────┘
     │           │
     └─────┬─────┘
           │
           ▼
    ┌──────────────────┐
    │ Sentiment        │
    │ Analysis Agent   │
    └─────────┬────────┘
              │
              ▼
       ┌──────────────┐
       │Store Results │
       │to Snowflake  │
       │(Append Mode) │
       └──────────────┘
```

**Note**: The fetch step automatically excludes comments already present in `COMMENT_SENTIMENT_FEATURES`, enabling incremental processing.

## Logging

Logs are automatically created in the `logs/` directory with timestamps:

```
logs/comment_processing_20251001_143022.log
```

## Adding New Data Sources

The system is designed to make adding new data sources easy:

### Steps to Add a New Source:

1. **Update Configuration** (`config_files/data_sources_config.json`):
   ```json
   "your_new_source": {
       "name": "Your New Source Name",
       "enabled": true,
       "sql_query_file": "sql/fetch_your_source.sql",
       "output_config": {
           "table_name": "YOUR_SOURCE_SENTIMENT_FEATURES",
           "database": "SOCIAL_MEDIA_DB",
           "schema": "ML_FEATURES"
       },
       "additional_fields": ["FIELD1", "FIELD2"]  // Optional
   }
   ```

2. **Create SQL Query File** (`sql/fetch_your_source.sql`):
   - Fetch comments with consistent column names
   - Include self-join for parent comments if available
   - Exclude already-processed comments (LEFT JOIN with output table)

3. **Create Table Initialization Script** (`sql/init_your_source_table.sql`):
   - Creates empty table structure
   - Base schema on `init_musora_table.sql`
   - Add source-specific fields as needed
   - **Run this in Snowflake FIRST before processing**

4. **Create Full Schema** (optional):
   - Base schema on `create_musora_ml_features_table.sql`
   - Include views and indexes

5. **Run First Time**:
   ```bash
   # Step 1: Run init script in Snowflake
   sql/init_your_source_table.sql

   # Step 2: Process first batch
   python main.py --overwrite --data-source your_new_source --limit 100
   ```

**No code changes required!**

## Best Practices

1. **Testing**: Always test with `--limit` flag first (e.g., `--limit 100`)
2. **New Data Sources**: Test new sources with `--sequential --limit 100` first
3. **Debugging**: Use `--sequential` flag for easier debugging of processing issues
4. **Incremental Processing**: Run regularly without `--overwrite` to process only new comments
5. **Monitoring**: Check logs for processing errors and batch completion
6. **Performance**: Use default parallel mode for production workloads
7. **Extensibility**: Follow the base agent pattern for consistency
8. **Error Handling**: All agents include robust error handling
9. **Failed Comments**: Review logs for failed comments - they'll be automatically retried in future runs
10. **Resource Management**: System automatically adapts to available CPU resources
11. **Parent Comments**: Ensure SQL queries include parent comment joins for best accuracy

## Sentiment Analysis Features

### Multi-Label Intent Classification

The system supports **multi-label intent classification**, meaning a single comment can have multiple intents:

- **Example**: "This is amazing! What scale are you using?" → `["praise", "question"]`
- **Example**: "Love this but can you make a tutorial on it?" → `["praise", "request"]`

### Context-Aware Analysis with Parent Comment Support

The sentiment analysis agent provides rich context understanding:

1. **Content Context**: Uses the `content_description` field to understand what the comment is about
2. **Parent Comment Context** (NEW): When analyzing reply comments, the system:
   - Automatically detects when a comment is a reply
   - Fetches the parent comment text from the database
   - Includes parent comment in the LLM prompt
   - Explicitly instructs the LLM that this is a reply comment
   - Results in more accurate sentiment and intent classification

**Example**:
- Parent Comment: "Does anyone know how to play this riff?"
- Reply Comment: "Yes!"
- Without parent context: Might be classified as unclear/off-topic
- With parent context: Correctly classified as answering a question

This dramatically improves accuracy for:
- Short reply comments ("Yes", "Thanks!", "Agreed")
- Sarcastic replies (context crucial for understanding)
- Continuation of discussions
- Agreement/disagreement comments

### Failure Handling & Reprocessing

Comments that fail sentiment analysis (missing critical fields like sentiment_polarity or intents) are:
- Marked as `success=False` in the workflow
- **NOT stored in Snowflake**
- **Automatically available for reprocessing** in future runs

This ensures only successfully processed comments are stored, while failed comments remain available for retry.

### Incremental Processing & Deduplication

The pipeline automatically handles incremental processing:
- **SQL-level deduplication**: Query excludes comments already in `COMMENT_SENTIMENT_FEATURES` using `LEFT JOIN`
- **Automatic retry**: Failed comments (not stored) are automatically retried on next run
- **Append-only mode**: Default behavior appends new records without overwriting
- **Production-ready**: Safe to run daily/weekly/monthly to process new comments

### Scalable Configuration

To add or modify sentiment categories or intents:

1. Edit `config_files/sentiment_analysis_config.json`
2. Add/modify categories in the `sentiment_polarity` or `intent` sections
3. Update `reply_policy.requires_reply_intents` if needed
4. No code changes required!

## Future Extensions

The modular architecture supports easy addition of:

- Topic classification agent
- Entity extraction agent
- Engagement score prediction agent
- Named entity recognition agent

Simply create a new agent inheriting from `BaseAgent` and add it to the workflow graph.

## Troubleshooting

### Issue: "Object does not exist or not authorized" on First Run

**Error**: `Object 'SOCIAL_MEDIA_DB.ML_FEATURES.MUSORA_COMMENT_SENTIMENT_FEATURES' does not exist or not authorized`

**Cause**: The fetch query tries to check for already-processed comments, but the output table doesn't exist yet on first run.

**Solution**:
1. Run the initialization script first:
   ```sql
   -- Execute in Snowflake
   sql/init_musora_table.sql
   ```
2. Then run the processing:
   ```bash
   python main.py --overwrite --data-source musora_comments --limit 100
   ```

### Issue: API Rate Limits

If hitting API rate limits, reduce the number of parallel workers or process fewer comments:
```bash
# Process fewer comments at a time
python main.py --limit 500

# Or use sequential mode
python main.py --sequential --limit 100
```

### Issue: Memory Issues

Process in smaller batches using `--limit`:
```bash
python main.py --limit 500
```

### Issue: Debugging Processing Errors

Use sequential mode to debug issues more easily:
```bash
python main.py --sequential --limit 50
```

This processes all comments in a single batch with clearer error messages.

### Issue: Connection Timeouts

Check Snowflake credentials in `.env` and network connectivity.

### Issue: Parallel Processing Not Working

If multiprocessing issues occur, use sequential mode:
```bash
python main.py --sequential
```

## Performance

### Expected Speedup

Parallel processing provides significant performance improvements:

- **Sequential**: 1x (baseline)
- **2 workers**: ~1.8-1.9x faster
- **5 workers**: ~4-4.5x faster

Speedup isn't perfectly linear due to:
- Snowflake connection overhead
- LLM API rate limits (shared across workers)
- I/O operations

### Monitoring Performance

The processing summary includes:
- Total processing time
- Average time per comment
- Number of workers used
- Batch size calculations
- Failed batches (if any)

## License

Internal use only - Musora sentiment analysis project.