Autonomous Blog Generation Agent

Turns a topic, a YouTube URL, or an uploaded audio/video file into a blog post, with optional translation. Generation logic is a LangGraph DAG, served through FastAPI, with a React frontend.

Pipeline

START
  ├─ (youtube) ─▶ fetch_transcript ─▶ brainstorm_titles ─▶ generate_content ─┬─ (target_language) ─▶ translate ─▶ END
  └─ (topic)   ───────────────────▶ brainstorm_titles ─▶ generate_content ─┘
                                                                            └─ (no translation) ─▶ END

• fetch_transcript — resolves a YouTube URL to a video ID and pulls its transcript (youtube-transcript-api), truncated to max_transcript_chars (default 8000).
• brainstorm_titles — Groq call at temperature=0.9, structured output, returns 3-5 candidate titles plus a selected best one.
• generate_content — Groq call at temperature=0.5, writes the full Markdown body for the selected title.
• translate — only runs if target_language is set and supported. Groq call at temperature=0.2, structured output, translates title + body in one shot while keeping the Markdown structure intact.

State is one shared TypedDict (app/graph/state.py) that nodes read from and partially update — raw_input, transcript, titles, selected_title, blog_content, target_language, translated_title/translated_content, plus error/error_stage.

Nodes don't raise on failure — they set error/error_stage on the state instead, and a routing function right after each node checks for that and jumps straight to END if something went wrong. LLM calls are wrapped so rate limits get their own error_stage (e.g. generate_content_rate_limited), separate from other failures — the API layer maps each stage to the right HTTP status (422/404/502/503) instead of a blanket 500.

Uploaded audio/video files skip the graph entirely: POST /transcribe-upload sends the file to Groq's hosted Whisper (whisper-large-v3-turbo), and the resulting transcript gets resubmitted to /generate as a normal topic input.

Stack

Backend: Python 3.12, FastAPI, LangGraph, LangChain (langchain-groq), Groq (llama-3.3-70b-versatile for text, whisper-large-v3-turbo for transcription), youtube-transcript-api, Pydantic / pydantic-settings, LangSmith (optional tracing), Pytest.

Frontend: React 18 + Vite, Tailwind CSS, react-markdown.

Project structure

app/
  api/routes.py          /generate, /transcribe-upload, /languages, /health
  config.py              Groq key/model, supported languages, LangSmith config
  graph/
    builder.py            DAG construction
    nodes.py              node implementations + routing
    state.py              shared graph state
  models/schemas.py       request/response models
  services/
    llm.py                 Groq chat model factory
    youtube.py              URL parsing + transcript fetching
    transcription.py        Whisper transcription
frontend/src/             App.jsx, api.js, components/
tests/                    offline test suite (mocked LLM + transcript fetch)

Setup

python3 -m venv .venv && source .venv/bin/activate
pip install -r requirements-dev.txt   # or requirements.txt for runtime-only
cp .env.example .env   # set GROQ_API_KEY (free tier: console.groq.com)

cd frontend && npm install
cp .env.example .env   # VITE_API_BASE_URL defaults to http://127.0.0.1:8000

Running

uvicorn app.main:app --reload      # http://127.0.0.1:8000/docs
cd frontend && npm run dev         # http://localhost:5173

API

curl -X POST http://127.0.0.1:8000/generate \
  -H "Content-Type: application/json" \
  -d '{"input_type": "topic", "content": "The history of coffee"}'

curl -X POST http://127.0.0.1:8000/generate \
  -H "Content-Type: application/json" \
  -d '{"input_type": "youtube", "content": "https://youtu.be/VIDEO_ID", "target_language": "french"}'

curl -X POST http://127.0.0.1:8000/transcribe-upload -F "file=@/path/to/clip.mp3"

GET /languages — supported translation targets. GET /health — liveness check.

Failure	Status
Malformed YouTube URL	422
Transcript disabled / not found	422
Video unavailable / private	404
YouTube blocked the request	503
Unsupported target_language	422
Empty / oversized file upload	422 / 413
Groq API error	502
Groq rate limit	503
Unexpected pipeline failure	500

Testing

pytest -q

Runs fully offline (mocked LLM + transcript fetch via tests/conftest.py). To hit the real Groq API and a real YouTube video, set GROQ_API_KEY and use /generate directly.