Automated academic research pipeline. Download a video from any of 1,750+ supported sites (YouTube, Vimeo, Twitch, TikTok, academic platforms, and more), transcribe it with word-level timestamps, search peer-reviewed literature across four academic databases, verify claims against the evidence, and generate a fully cited research report — all with one command.

---

What It Does

YouTube Research AI transforms passive video watching into active academic research. It's not a summarizer — it's a full research pipeline that treats a video as a primary source and builds a cited, verified research report around it.

The Pipeline

Video URL → Download (yt-dlp) → Extract Audio (ffmpeg) → Transcribe (Whisper) 
  → Domain Detection → Literature Search (arXiv, Semantic Scholar, CrossRef, OpenAlex) 
  → Evidence Synthesis → Claim Verification → Deep Research Iterations 
  → Cited Report Generation → Reference Formatting

Output Per Video

Output File	Contents
transcripts/<slug>.transcript.json	Full transcription with word-level timestamps
transcripts/<slug>.transcript.txt	Plain text transcript
transcripts/<slug>.timestamped.md	Timestamped transcript with [HH:MM:SS] markers
reports/<slug>.research.json	Structured research data (JSON)
reports/<slug>.research.md	Polished standalone research article with citations
research/search-queries.jsonl	Audit log of every academic search query
research/sources.json	All retrieved and selected sources
synthesis/synthesis.md	Cross-source synthesis (multi-video mode)

What the Report Contains

• Executive summary with key findings
• Core thesis extracted from the material
• Detailed summary of all substantive content
• Key ideas with explanations and significance
• Timeline of important moments with timestamps
• Glossary of technical terms and definitions
• Memorable quotes preserved verbatim
• Claims to verify — flagging assertions needing evidence
• Peer-reviewed citations — inline numbered references to academic sources
• Evidence quality assessment — confidence level per claim (well-supported, plausible, contested, speculative, opinion)
• Source quality rating — overall epistemic quality of the source material
• Literature gap analysis — what the academic literature covers that the video misses
• Research methods section — describing how the automated research was conducted
• Formatted reference list — APA, Chicago, or IEEE style
• Study questions and practical applications
• Follow-up research recommendations
• 7-day study plan for deep learning

---

Quick Start

Prerequisites

• Node.js 22+ — Download
• yt-dlp — brew install yt-dlp (macOS) or pip install yt-dlp
• ffmpeg — brew install ffmpeg (macOS) or apt install ffmpeg
• An AI provider API key — OpenAI, Anthropic, Google, or OpenAI-compatible

Install

git clone https://github.com/jongan69/YouTubeResearchAI.git
cd YouTubeResearchAI
npm install
cp .env.example .env

Configure

Edit .env with your AI provider API key:

# OpenAI (recommended for full feature set)
AI_PROVIDER=openai
OPENAI_API_KEY=sk-your-key-here

# Or Anthropic Claude
# AI_PROVIDER=anthropic
# ANTHROPIC_API_KEY=sk-ant-your-key-here
# OPENAI_API_KEY=sk-your-key-here  # needed for Whisper transcription

# Or Google Gemini
# AI_PROVIDER=google
# GOOGLE_API_KEY=your-key-here
# OPENAI_API_KEY=sk-your-key-here  # needed for Whisper transcription

# Or any OpenAI-compatible endpoint (Groq, DeepSeek, Ollama, etc.)
# AI_PROVIDER=openai-compat
# OPENAI_API_KEY=your-key-here
# OPENAI_BASE_URL=https://api.groq.com/openai/v1

Run

# Add video URLs to links.txt (one per line — any site yt-dlp supports)
echo "https://www.youtube.com/watch?v=VIDEO_ID" > links.txt

# Basic run — transcript + study-guide report
npm run research

# PhD-grade research — full literature search + citations + evidence verification
npm run research -- --research --verify

# Maximum depth — deep iterative research with all features
npm run research -- --research --research-depth deep --verify --citation-style apa --max-sources 20

Or double-click RUN.command on macOS.

Verify Setup

npm run doctor

---

Features

🔬 Automated Literature Research

The system searches four free academic databases simultaneously for every claim in the video:

Database	Coverage	Authentication
arXiv	CS, physics, math, statistics preprints	None (polite pool)
Semantic Scholar	200M+ papers across all disciplines	Free API key (optional, raises rate limit)
CrossRef	150M+ peer-reviewed journal articles, proceedings, chapters	None (polite pool with mailto)
OpenAlex	250M+ works, broadest open-access coverage	None

How it works:

1. Query planning — The LLM extracts 3–6 researchable topics from the transcript and generates 2–3 precise academic search queries per topic
2. Parallel search — Queries are sent to all enabled databases with controlled concurrency
3. Deduplication and scoring — Results are merged by DOI, scored by citation count + recency + keyword relevance
4. Source injection — Selected papers are injected into the report prompt so the LLM can cite them inline
5. Citation validation — Inline [S1]…[Sn] markers are validated against the fetched metadata; hallucinated citations are pruned
6. Reference formatting — A formatted bibliography is appended in APA, Chicago, or IEEE style

🧪 Evidence Synthesis & Claim Verification

With --verify, each claim extracted from the video is compared against the retrieved literature:

• Well-supported ✅ — Multiple papers agree
• Plausible 🟡 — Consistent with literature but not directly tested
• Contested ⚠️ — Literature shows genuine disagreement
• Speculative ❓ — No evidence found in retrieved sources
• Opinion 💬 — Normative claim, not empirically verifiable

Supporting and contradicting sources are listed per claim, and an Evidence Quality Assessment section is appended to the report.

🎓 Domain-Specific Research

Five academic domain profiles automatically tailor the research strategy:

Domain	Detection Triggers	Preferred APIs	Evaluation Standards
Computer Science	algorithm, API, Rust, React, ML, GPU…	arXiv, Semantic Scholar	Peer-reviewed proceedings, benchmark rigor, reproducibility
Medicine	clinical, diagnosis, RCT, trial, drug…	Semantic Scholar, CrossRef	Evidence hierarchy, sample size, conflicts of interest
Social Sciences	economics, psychology, policy, survey…	CrossRef, Semantic Scholar	Causal identification, sample representativeness, replication
Humanities	philosophy, history, ethics, discourse…	CrossRef, OpenAlex	Primary source engagement, interpretive framework awareness
Natural Sciences	physics, chemistry, biology, genetics…	arXiv, Semantic Scholar	Experimental design, measurement precision, replication status

Domain detection is automatic based on keyword density in the transcript. Force a domain with --domain computer-science.

🔄 Deep Iterative Research

--research-depth deep enables multi-pass research:

1. Initial search — standard literature retrieval
2. Gap analysis — identify low-confidence claims and literature gaps
3. Targeted re-search — new queries for gaps
4. Re-verification — expanded evidence synthesis with all sources
5. Convergence — stops when no new sources are found or max iterations reached

📊 Multi-Source Synthesis

With --synthesis and 2+ videos in links.txt, the system generates a cross-source synthesis report identifying:

• Consensus themes — where sources agree
• Contradictions — where sources disagree
• Literature gaps — important topics no source covers
• Unique contributions — what each source adds

🖼️ Visual Content Analysis

With --vision, keyframes are extracted from the video and analyzed by vision-capable AI:

• Equations — reproduced in LaTeX where possible
• Diagrams — described structurally
• Code — language and key operations identified
• Charts and tables — data extracted and contextualized
• On-screen citations — transcribed exactly

Requires a vision-capable model (GPT-5.5, Claude, Gemini). Automatically skipped for providers without vision support.

🏗️ Provider-Agnostic Architecture

All four AI providers are supported with feature parity where the provider allows:

Feature	OpenAI	Anthropic	Google	OpenAI-compat
Report generation	✅ GPT-5.5	✅ Claude	✅ Gemini	✅ Groq, DeepSeek, Ollama
Transcription	✅ Whisper	❌ (needs OpenAI key)	❌ (needs OpenAI key)	✅ (if supported)
Reasoning/thinking	✅ reasoning.effort	✅ thinking.budget_tokens	✅ thinkingConfig	✅ System prompt guidance
Structured output	✅ json_schema strict	✅ Tool use + tool_choice	✅ response_schema	✅ json_schema / json_object fallback
Vision (Phase 5)	✅ image_url base64	✅ image content blocks	🚧 Planned	❌ (env override available)
Zero new dependencies	✅ openai SDK only	✅ fetch() only	✅ fetch() only	✅ openai SDK only

---

Complete CLI Reference

Core Options

Flag	Env Var	Default	Description
--links FILE	—	./links.txt	File with video URLs, one per line (any yt-dlp supported site)
--out-dir DIR	—	./outputs	Output root directory
--run-name NAME	—	run-YYYY-MM-DD-HHMMSS	Custom run folder name
--ai-provider ID	AI_PROVIDER	openai	openai, anthropic, google, openai-compat
--transcription-model ID	OPENAI_TRANSCRIPTION_MODEL	whisper-1	Transcription model
--report-model ID	—	Provider default	Report generation model
--reasoning-effort LEVEL	OPENAI_REASONING_EFFORT	medium	low, medium, high
--verbosity LEVEL	OPENAI_TEXT_VERBOSITY	medium	low, medium, high
--max-output-tokens N	—	Provider default	Report output token budget
--transcript FILE	—	—	Generate report from existing transcript
--title TEXT	—	—	Report title (with --transcript)
--skip-download	—	—	Treat links as local video paths
--no-report	—	—	Download and transcribe only

Research Options

Flag	Env Var	Default	Description
--research	RESEARCH_ENABLED	off	Enable literature search + citations
--research-depth LEVEL	RESEARCH_DEPTH	medium	none, light, medium, deep
--research-topics "a; b"	RESEARCH_TOPICS	auto	Semicolon-separated topics
--max-sources N	RESEARCH_MAX_SOURCES	10	Sources to cite in report
--max-papers-per-topic N	RESEARCH_MAX_PAPERS_PER_TOPIC	5	Papers per query per source
--research-apis LIST	RESEARCH_APIS	all four	arxiv,semantic_scholar,crossref,openalex
--citation-style STYLE	CITATION_STYLE	apa	apa, chicago, ieee
--verify	VERIFY_ENABLED	off	Evidence synthesis + claim verification
--domain ID	DOMAIN	auto	Force domain profile
--synthesis	SYNTHESIS	off	Cross-source synthesis (2+ videos)
--vision	VISION_ENABLED	off	Visual frame analysis
--max-frames N	MAX_FRAMES	20	Max frames to extract
--research-iterations N	RESEARCH_ITERATIONS	3	Max deep research iterations

Environment Variables

Variable	Required	Description
AI_PROVIDER	No (default: openai)	openai, anthropic, google, openai-compat
OPENAI_API_KEY	For OpenAI / transcription	OpenAI API key
OPENAI_BASE_URL	For openai-compat	Compatible endpoint URL
ANTHROPIC_API_KEY	For Anthropic	Anthropic API key
GOOGLE_API_KEY	For Google	Google Gemini API key
SEMANTIC_SCHOLAR_API_KEY	No (optional)	Raises rate limit from 100 to 1000 req/5min
RESEARCH_MAILTO	No (recommended)	CrossRef polite pool identification
CITATION_STYLE	No (default: apa)	Citation formatting style

---

Examples

Basic: Study Guide from a Lecture

npm run research

Generates a polished standalone article with executive summary, key ideas, glossary, study questions, and a 7-day study plan.

Research: Cited Report with Peer-Reviewed Sources

npm run research -- --research --citation-style apa --max-sources 10

Searches arXiv, Semantic Scholar, CrossRef, and OpenAlex for relevant papers, injects 10 best sources, and generates a cited report with APA-formatted reference list.

Deep: PhD-Grade Research with Evidence Verification

npm run research -- \
  --research \
  --research-depth deep \
  --verify \
  --citation-style apa \
  --max-sources 20 \
  --reasoning-effort high \
  --verbosity high \
  --max-output-tokens 32000

Multi-pass iterative research with evidence synthesis, confidence levels per claim, and maximum academic rigor.

Multi-Source: Synthesize Across Multiple Videos

# links.txt contains 2+ related videos
npm run research -- --research --verify --synthesis

Generates individual cited reports plus a cross-source synthesis identifying consensus, disagreements, and gaps.

Visual: Analyze Slides, Code, and Equations

npm run research -- --research --vision --max-frames 30

Extracts keyframes from the video, analyzes visual content (equations, diagrams, code, charts), and integrates findings into the report.

Resume: Report from Saved Transcript

npm run research -- \
  --transcript "outputs/run-2026-08-04-065333/transcripts/my-video.timestamped.md" \
  --title "My Research Topic" \
  --research --verify

Reuses an existing transcript without re-downloading or re-transcribing.

Provider-Specific

# Anthropic Claude with high thinking budget
npm run research -- --ai-provider anthropic --report-model claude-opus-5 --reasoning-effort high

# Google Gemini
npm run research -- --ai-provider google --report-model gemini-2.5-pro

# Groq (fast Llama)
npm run research -- --ai-provider openai-compat --report-model llama-3.3-70b-versatile

---

Output Structure

outputs/run-YYYY-MM-DD-HHMMSS/
├── links.txt                          # Copy of input links
├── manifest.json                      # Full run metadata
├── downloads/
│   └── Video Title [videoId].mp4      # Downloaded video + .info.json
├── transcripts/
│   ├── video-slug.transcript.json     # Whisper verbose_json (word timestamps)
│   ├── video-slug.transcript.txt      # Plain text
│   └── video-slug.timestamped.md      # Timestamped markdown
├── reports/
│   ├── video-slug.research.json       # Structured research data
│   └── video-slug.research.md         # Polished standalone article
├── research/                          # (--research enabled)
│   ├── search-queries.jsonl           # Every search query logged
│   └── sources.json                   # Retrieved + selected sources
├── frames/                            # (--vision enabled)
│   └── video-slug/
│       ├── frame-001.jpg
│       └── frame-002.jpg
└── synthesis/                         # (--synthesis enabled)
    ├── synthesis.json
    └── synthesis.md

---

Architecture

scripts/
├── process-links.mjs          # Pipeline conductor
├── ai-config.mjs              # Configuration resolution (env + CLI)
├── doctor.mjs                 # Environment validation
├── lib.mjs                    # Utilities (slugify, timestamp, args parser)
├── prompts/
│   └── index.mjs              # All LLM schemas and prompt builders
├── domains/
│   └── index.mjs              # 5 domain profiles with auto-detection
├── providers/
│   ├── interface.mjs          # Provider factory + retry logic
│   ├── openai.mjs             # OpenAI (Responses API + vision)
│   ├── anthropic.mjs          # Anthropic (Messages API + thinking + vision)
│   ├── google.mjs             # Google (Gemini API)
│   └── openai-compat.mjs      # OpenAI-compatible (Groq, DeepSeek, Ollama)
└── research/
    ├── index.mjs              # Research orchestrator
    ├── citation-manager.mjs   # Reference tracking + formatting
    ├── literature-search.mjs  # 4 academic API adapters
    ├── evidence-synthesis.mjs # Claim verification engine
    ├── iterative-research.mjs # Deep multi-pass research
    ├── synthesis.mjs          # Cross-source synthesis
    ├── vision-analysis.mjs    # Frame extraction + vision AI
    └── provenance.mjs         # JSONL audit logging

Design principles:

• Zero new npm dependencies — everything uses Node.js built-in fetch() and the existing openai SDK
• All external APIs are free — arXiv, Semantic Scholar, CrossRef, OpenAlex require no payment
• Provider-agnostic — every LLM feature degrades gracefully when unsupported
• Strictly additive — default behavior is byte-compatible with the original tool
• Graceful degradation — any optional feature failing does not crash the pipeline

---

Cost Estimates

Mode	Transcription	LLM Passes	Research APIs	Approximate Cost (OpenAI)
Basic (no flags)	1× Whisper	1× GPT-5.5	None	~$0.10–0.30/video
Research (--research)	1× Whisper	2× GPT-5.5 (query plan + report)	15–30 free API calls	~$0.20–0.60/video
Research + Verify (--research --verify)	1× Whisper	3× GPT-5.5	15–30 free API calls	~$0.30–0.90/video
Deep (--research-depth deep)	1× Whisper	5–8× GPT-5.5	50–100 free API calls	~$0.50–2.00/video

All academic database queries are free. Costs only come from LLM API calls.

---

FAQ

Does this replace peer review? No. The system retrieves and cites peer-reviewed sources, but it does not perform peer review itself. It's a research assistant, not a replacement for expert judgment.

How accurate are the citations? Citations are validated against fetched metadata (DOI, title, authors). The system prunes hallucinated citations that don't match any retrieved source. However, the LLM may still miscorrelate a claim with a source — the evidence synthesis pass helps catch this.

What if an academic API is down? The pipeline continues without that source. Failures are logged to the audit trail and reported in the console.

What video sources are supported? Any of the 1,750+ sites that yt-dlp supports — YouTube, Vimeo, Twitch, TikTok, Twitter/X, Facebook, Instagram, Dailymotion, Bilibili, Coursera, edX, academic lecture platforms, podcasts, and more. The pipeline auto-detects the source and applies the right download strategy. With --skip-download, any local video file works regardless of origin.

What languages does transcription support? Whisper supports 99 languages. Transcription quality varies by language.

Is my data private? All processing is local except for API calls (transcription to OpenAI, report generation to your chosen AI provider, literature search to public academic databases). No data is stored on external servers beyond what those APIs normally log.

---

Contributing

Contributions welcome. Areas of interest:

• Additional academic database adapters (PubMed, Scopus, Web of Science)
• Additional domain profiles
• Export formats (LaTeX, docx, PDF)
• Web UI / Electron app
• Docker containerization
• Additional citation styles (MLA, Harvard, Vancouver)
• Full-text retrieval and analysis (beyond metadata)

See CONTRIBUTING.md for guidelines.

---

License

MIT — see LICENSE for details.

---

YouTube Research AI — From passive watching to active research.
Built with ❤️ for lifelong learners, researchers, and the insatiably curious.