Call Analysis Backend

A FastAPI-based backend service for analyzing call transcripts using Google's Gemini AI.

Features

• Transcript storage and retrieval
• AI-powered call analysis using Google Gemini 2.0 Flash-Lite
• Structured data extraction from call transcripts
• JSON data storage in PostgreSQL
• RESTful API for integration with frontend applications

Prerequisites

• Python 3.12+
• Google Gemini API key
• uv >=0.6.9
• Postgres >=17

Setup

1. Clone the repository

git clone git@github.com:ai-Posibl/mono_backend.git

2. Use uv to setup python venv and install packages

uv sync

3. Activate the virtual environment

# For Windows
.venv\Scripts\activate
# For Linux
source .venv/bin/activate 

3. Set up environment variables

Create .env file with reference to env.example

6. Run the application

uvicorn main:app --reload

The API will be available at http://localhost:8000 The docs will be presetn at http://localhost:8000

API Endpoints

Call Analysis

• POST /analyze

  • Analyzes a call transcript using Gemini AI
  • Request body:

    { 
      "session_id": "string",
      "transcript": "string",
      "audio_url": "string"
    }

  • Response:

    {
      "session_id": "string",
      "transcript": "string",
      "summary": "string",
      "structured_output": {
        "business_name": "string",
        "business_offering": "string",
        "transactions_per_month": "string"
      },
      "success_evaluation": {
        "result": true
      },
      "audio_url": "string"
    }

• GET /analysis/{session_id}

  • Retrieves a previously analyzed call by session ID
  • Response: Same as the analyze endpoint response

• POST /add

  • Adds a new transcript to the database
  • Request body:

    { 
      "session_id": "string",
      "transcript": "string"
    }

• GET /all

  • Retrieves all transcripts from the database

Project Structure

backend/
├── app/
│   ├── models/
│   │   ├── __init__.py
│   │   ├── audio_model.py
│   │   ├── transcript.py
│   │   └── callAnalysis_model.py
│   ├── routes/
│   │   └── callAnalysis_router.py
│   ├── database.py
│   ├── dependencies.py
│   └── main.py
├── .env
├── .gitignore
└── README.md

Development

Adding New Features

1. Create new models in the app/models directory
2. Add new routes in the app/routes directory
3. Update the main application in app/main.py

Database Schema

The application uses PostgreSQL with the following tables:

• transcripts: Stores raw call transcripts
• audio: Stores audio file information
• call_analysis: Stores call analysis results including JSON data

Working with JSON in PostgreSQL

This application uses PostgreSQL's JSON data type for storing structured data:

• structured_output: Stores business information extracted from the call
• success_evaluation: Stores the evaluation result of the call

To query JSON fields in PostgreSQL, you can use operators like:

-- Example: Find all successful calls
SELECT * FROM call_analysis WHERE success_evaluation->>'result' = 'true';

-- Example: Find calls for a specific business type
SELECT * FROM call_analysis WHERE structured_output->>'business_offering' LIKE '%retail%';

Common Issues

Gemini API Schema Limitations

When using structured output with Gemini API:

• Union types (e.g., int | str) are not supported in response schemas
• Always use single types for each field (e.g., str, int, bool)
• For numeric values that might be strings, use str type and convert as needed