Case Study

YouTube Research Agent (2025) — Automating YouTube Surveying with LLM Summaries

An open-source Python CLI tool that searches YouTube for high-view-count videos on a topic, extracts transcripts, summarizes each video using GPT-3.5 Turbo via a LangChain map-reduce chain, and emails a formatted digest to a recipient.

Tech Stack

Problem

Researchers and knowledge workers often need to survey YouTube for information on a topic, which requires manual searching, skimming, and synthesizing notes across multiple videos.

Solution

Given a topic and recipient email, the agent:

1. 1Searches YouTube for relevant, high-view-count videos
2. 2Extracts transcripts
3. 3Summarizes each video using GPT-3.5 Turbo with a map-reduce summarization chain
4. 4Emails a digest of the top videos and their summaries to a specified recipient

Architecture & Data Flow

• CLI collects: topic, recipient email, max videos (default 5)
• Search module queries YouTube Data API v3, filters to videos published after 2024-01-01, orders by view count, fetches transcripts per video (handles missing transcripts)
• Summarizer combines title + description + transcript, splits text into 2000-character chunks with 200-character overlap using RecursiveCharacterTextSplitter, runs LangChain map-reduce summarization using ChatOpenAI (gpt-3.5-turbo, temperature=0)
• Email module formats summaries into plain-text and sends via SMTP/TLS (default Gmail SMTP: port 587)
• Output: returns summaries to CLI for console display and sends the email digest

Key Files

• src/main.py— Orchestrator (YouTubeResearchAgent) + CLI input
• src/config/settings.py— loads env vars (YouTube/OpenAI/email/SMTP)
• src/youtube/search.py— search + transcript extraction
• src/summarizer/summarize.py— chunking + map-reduce summarization chain
• src/email/sender.py— formats and sends email via SMTP/TLS
• tests/— present but currently empty placeholders

Design Patterns

• Separation of concerns across search, summarization, email modules
• Orchestrator pattern in main.py
• Map-reduce summarization to handle long transcripts
• Centralized configuration loading via settings.py

Technical Architecture

A Python CLI tool with clear separation of concerns. The orchestrator (src/main.py) coordinates three modules: a search module (src/youtube/search.py) that queries YouTube Data API v3, filters to videos published after 2024-01-01, orders by view count, and extracts transcripts via youtube-transcript-api; a summarizer module (src/summarizer/summarize.py) that combines title + description + transcript, chunks text into 2000-character segments with 200-character overlap using RecursiveCharacterTextSplitter, and runs a LangChain map-reduce summarization chain using ChatOpenAI (gpt-3.5-turbo, temperature=0); and an email module (src/email/sender.py) that formats summaries into plain-text and sends via SMTP/TLS (default Gmail SMTP: port 587). Configuration is centralized in src/config/settings.py using python-decouple and python-dotenv.

Key Decisions

• 1

  Used LangChain map-reduce chain over simple prompt truncation — preserves full transcript content regardless of length

• 2

  Separated search, summarization, and email into distinct modules — clear boundaries make each component independently testable

• 3

  Used RecursiveCharacterTextSplitter with 200-character overlap — maintains context across chunk boundaries for coherent summaries

• 4

  Chose gpt-3.5-turbo with temperature=0 — deterministic output for research summarization, lower cost than GPT-4

• 5

  Centralized all config in settings.py via python-decouple — single source of truth for API keys and SMTP credentials

Tradeoffs

• CLI-only interface limits accessibility but keeps scope focused and avoids web infrastructure complexity
• Synchronous processing means longer wait times for multiple videos, but simplifies implementation and debugging
• Plain-text email limits formatting but avoids HTML rendering complexity and email client compatibility issues
• Hardcoded 2024-01-01 date filter limits flexibility but scopes results to recent content by default
• Stateless design (no caching/db) means repeated queries re-fetch and re-summarize, but avoids storage management

Current Limitations

• CLI-only (no web UI / API)
• Tests are stubs (empty)
• Uses print statements instead of structured logging
• Synchronous processing (no async/parallel)
• Hardcoded filter: videos after 2024-01-01
• No retry/rate limiting
• Plain-text email only
• Stateless (no caching/db)