Transcript Summarizer API

A lightweight, production-style REST API that uses an LLM to automatically analyze customer support transcripts.

Instead of a human reading thousands of conversations manually, this tool processes them instantly and returns structured, actionable data — regardless of transcript length.

---

What It Does

Send a support transcript to the API and receive:

Field	Description
summary	A concise summary of the conversation
sentiment	Overall customer sentiment (positive, neutral, negative)
risk_level	Risk of churn or escalation (low, medium, high)
action_items	List of specific follow-up actions identified
was_chunked	Whether the transcript was split into chunks for processing
chunks_processed	Number of chunks the transcript was split into

---

Transcript Chunking Pipeline

LLMs have context window limits. Long transcripts become expensive, reduce output quality, and can exceed those limits entirely.

This API solves that with an automatic chunking pipeline:

Transcript
   ↓
Split into overlapping chunks (default: 500 words, 50-word overlap)
   ↓
Analyze all chunks in parallel (asyncio.gather)
   ↓
Merge results:
  - Summary    → LLM combines partial summaries into one
  - Sentiment  → worst-case across all chunks
  - Risk level → highest risk across all chunks
  - Actions    → deduplicated union across all chunks

Short transcripts skip chunking entirely and go through a single LLM call. The was_chunked field in the response tells you which path was taken.

Chunk analysis runs with asyncio.gather — all chunks are sent to the OpenAI API simultaneously, so total latency equals the slowest single chunk rather than the sum of all of them.

---

Tech Stack

• Python — core language
• FastAPI — REST API framework
• OpenAI API — GPT-4o-mini for transcript analysis
• Pydantic — request/response validation and typed models
• Uvicorn — ASGI server

---

Project Structure

Transcript Summarizer API/
├── main.py           # FastAPI app — defines routes
├── analyzer.py       # OpenAI integration, chunking orchestration, merge logic
├── chunker.py        # Word-based transcript splitter with overlap
├── models.py         # Pydantic models for request/response
├── requirements.txt  # Python dependencies
├── .env.example      # Environment variable template
└── .gitignore

---

Getting Started

1. Clone the repository

git clone https://github.com/guillermolopezareal/transcript-summarizer-api.git
cd transcript-summarizer-api

2. Create and activate a virtual environment

python -m venv venv

# Windows
.\venv\Scripts\Activate.ps1

# Mac/Linux
source venv/bin/activate

3. Install dependencies

pip install -r requirements.txt

4. Set up your environment variables

cp .env.example .env

Open .env and add your OpenAI API key:

OPENAI_API_KEY=sk-...your key here...

You can get an API key at https://platform.openai.com/api-keys

5. Start the server

uvicorn main:app --reload

The API will be running at http://localhost:8000.

---

API Endpoints

GET /health

Health check — confirms the server is running.

Response:

{ "status": "ok" }

---

POST /analyze

Analyzes a customer support transcript. Automatically chunks long transcripts.

Request body:

{
  "transcript": "Agent: Hello, how can I help you today?...",
  "chunk_size_words": 500
}

Field	Type	Default	Description
transcript	string	required	The full text of the support transcript
chunk_size_words	integer	500	Max words per chunk (range: 10–2000)

Response:

{
  "summary": "The customer contacted support regarding a delayed order and a duplicate charge. Both issues were resolved — a replacement was shipped and a refund was issued.",
  "sentiment": "negative",
  "risk_level": "high",
  "action_items": [
    "Send express replacement with tracking number",
    "Process refund for duplicate charge within 3-5 business days"
  ],
  "was_chunked": true,
  "chunks_processed": 3
}

---

Example Tests

PowerShell — short transcript (no chunking)

$body = @{
    transcript = "Agent: Hello how can I help you today? Customer: My reimbursement has been delayed for two weeks and I want to cancel my policy. Agent: I apologize. I will escalate this immediately."
} | ConvertTo-Json

Invoke-RestMethod -Uri "http://localhost:8000/analyze" -Method Post -ContentType "application/json" -Body $body

PowerShell — force chunking

$body = @{
    chunk_size_words = 30
    transcript = "Agent: Hello how can I help you today? Customer: My reimbursement has been delayed for two weeks and I want to cancel my policy. Agent: I apologize. I will escalate this immediately and follow up within 24 hours. Customer: Also I was charged twice this month. Agent: I can see that duplicate charge and will issue a refund within 3 to 5 business days. Customer: Thank you. Agent: Is there anything else I can help with? Customer: No that is all. Agent: Have a great day."
} | ConvertTo-Json

Invoke-RestMethod -Uri "http://localhost:8000/analyze" -Method Post -ContentType "application/json" -Body $body

---

Interactive Docs

FastAPI generates automatic documentation at:

• Swagger UI: http://localhost:8000/docs
• ReDoc: http://localhost:8000/redoc

---

Real-World Use Cases

• Insurance claim call analysis
• Bank support quality assurance
• Technical support escalation detection
• CRM integration for churn risk flagging