---
title: CallCenterSummarizationAgent
emoji: 🚀
colorFrom: red
colorTo: red
sdk: docker
app_port: 8501
tags:
- streamlit
pinned: false
short_description: Summarize Call Center Conversations
license: mit
# Add arguments here if the SDK supports them:
args: ["--server.enableCORS", "false", "--server.enableXsrfProtection", "false"]
---
---
# Call Center Data Analysis Agent

This project implements a multi-agent workflow using **LangGraph** to process, transcribe, summarize, and score call center data. It comes with a **Streamlit** user interface to easily upload `.csv`, `.json`, `.mp3`, or `.wav` files and view the resulting insights.

## Workflow Flow & Agent Classes

The core analysis logic is driven by a LangGraph StateGraph defined in `src/workflow.py`. The state moves sequentially between several agent classes located in the `src/agents/` directory.

### Workflow Architecture
```text
                  [ IntakeAgent ]
                        |
                        v
               [ Router (file type) ]
              /         |           \
 (CSV invalid)/     (If Audio)     (If Text)
          v          v               v
        (END) [ TranscriptionAgent ] |
                  |                  |
                  v                  |
            [ ModerationAgent ] <-----/
                  |
                  v
          [ SummarizationAgent ]
                  |
                  v
     [ PostSummarizeRouter (output) ]
              |                |
           (END)        [ QualityScoringAgent ]
                               |
                               v
                             (END)
```

Notes:
- The workflow uses a LangGraph memory checkpointer (`MemorySaver`) and fallbacks on critical nodes (summarization/scoring) to avoid UI breakage on transient API/parse errors.
- If CSV headers are invalid, the workflow terminates immediately and the UI shows a validation error.
- After summarization, the workflow can short-circuit to `END` if the transcript is too short or the model output is missing/empty.

### Agent Classes

1. **`IntakeAgent` (`src/agents/IntakeAgent.py`)**
   - *Entry Point*.
   - Reads the uploaded file, validates the file format and schema, extracts basic metadata, and runs a first-pass clean-up (using an LLM).
   - **CSV requirement:** headers must include `id` and `transcript` (case-insensitive).
   - **JSON supported shapes:** a list of `{id, transcript}` objects, a single `{id, transcript}` object, or a dict with `transcripts`/`calls` arrays containing `{id, transcript}` objects.

2. **`Router` (`src/agents/Router.py`)**
   - *Conditional Routing Node*.
   - Determines the next step based on the file type and intake validation state.
   - **If Audio (`.mp3`, `.wav`)**: Routes to the `TranscriptionAgent`.
   - **If Text (`.csv`)**: Routes to the `ModerationAgent` then summarization/scoring.
   - **If CSV invalid**: Routes to `END` (the UI displays `metadata.intake_error`).

   **`PostSummarizeRouter` (`src/agents/Router.py`)**
   - Routes based on model output quality (e.g., short transcript or missing summary can skip scoring).

3. **`TranscriptionAgent` (`src/agents/TranscriptionAgent.py`)**
   - Utilizes `openai-whisper` to convert audio files into text.
   - Also scrubs the resulting transcript of profanity before passing it down the pipeline.

4. **`ModerationAgent` (`src/agents/ModerationAgent.py`)**
   - Receives text either directly from the `Router` (if text upload) or from the `TranscriptionAgent` (if audio upload).
   - Identifies any obscene words or profanity using an LLM and replaces them entirely with a `***` mask to safely prepare the text for downstream analysis.

5. **`SummarizationAgent` (`src/agents/SummarizationAgent.py`)**
   - Takes the redacted text from the `ModerationAgent`.
   - Generates a concise **summary**, **key points**, **action items**, **tags**, and **highlights** using OpenAI (`gpt-4o`) with Pydantic-structured output.

6. **`QualityScoringAgent` (`src/agents/QualityScoringAgent.py`)**
   - Takes the clean text and evaluates it against a predefined rubric.
   - Scores the transcript based on Tone, Professionalism, and Structured Resolution using Pydantic structured output (function calling when supported). Automatically applies a 3-point penalty to each score and logs a count of policy violations if the `ModerationAgent` detected and masked any profanity (`***`).

## Prerequisites

- Python 3.9+
- An OpenAI API key
- `ffmpeg` installed on your system (required for `openai-whisper` audio transcription). 
  - On macOS: `brew install ffmpeg`
  - On Ubuntu/Debian: `sudo apt update && sudo apt install ffmpeg`

## Installation

1. Create a virtual environment and activate it (if you haven't already):
   ```bash
   python3 -m venv .venv
   source .venv/bin/activate
   ```
2. Install the required dependencies:
   ```bash
   pip install -r requirements.txt
   ```

## Running the Application

1. Ensure your OpenAI API key is set in your environment variables:
   ```bash
   export OPENAI_API_KEY="your_api_key_here"
   ```

2. Start the Streamlit application:
   ```bash
   streamlit run src/streamlit_app.py
   ```

3. Open your browser to the local URL provided by Streamlit (usually `http://localhost:8501`).
4. Use the sidebar to upload a `.csv`, `.mp3`, or `.wav` file and watch the agents analyze your data!
 
## Processed Files
 
**Note:** The following sample data can be used for analysis:
- [Customer Call Center Dataset Analysis](https://www.kaggle.com/datasets/rafaqatkhan608/customer-call-center-dataset-analysis/code/data)
- [E-commerce Customer Support English Audio](https://huggingface.co/datasets/HumynLabs/e-commerce-customersupport-english-audio/tree/main)