🧠 Agent YouTube Journalism

🔍 Project overview

Agent YouTube Journalism is an open-source investigative AI assistant that transcribes, summarizes, analyzes, and answers questions about videos from YouTube — especially those related to Brazilian politics and public interest.

The system uses multi-agent reasoning and Retrieval-Augmented Generation (RAG) to:

• Transcribe YouTube videos in Brazilian Portuguese (using OpenAI Whisper API)
• Summarize the transcript with DeepSeek via Groq Cloud
• Search the web for context (via DuckDuckGo)
• Highlight journalistically relevant parts of the video
• Index and answer questions based on the transcript + general knowledge if needed

🟢 Try it online: https://agentytjournalism.streamlit.app/
⚠️ You must provide your own API keys for Groq, OpenAI, and Hugging Face.

---

⚙️ How it works

1. User enters YouTube video URL + API keys
2. The app:
   • Downloads audio via yt-dlp
   • Transcribes using Whisper (openai.Audio.transcribe)
   • Summarizes with Groq LLM (DeepSeek)
   • Searches the web for context
   • Highlights journalistic investigation leads
   • Indexes the transcript with FAISS
3. The user can:
   • View the analysis
   • Ask questions based on the video (with RAG + LLM knowledge fallback)

The system uses smolagents to structure the reasoning with a clear cycle:

Thought → Code → Observation

---

🗂️ Project structure

🔹 Main app

• app.py: Main Streamlit app with two tabs: Analysis & Questions
• process_video.py: Orchestrates full pipeline (transcription → summary → highlight → indexing)
• rag_question_tab.py: Handles the RAG-based Q&A flow with session state
• agent_config.py: Defines tools and setup for smolagents agent
• streamlit_app.yaml: Config file for deployment (Streamlit Community Cloud)
• prompts.yaml: Prompt templates used for summarization, analysis, and code reasoning

🔹 Groq integration

• groq_model.py: Executes prompts with Groq LLM and truncates long prompts when needed
• list_groq_models.py: Lists all Groq-hosted models available for querying

🔹 Tools (used by agents)

• tools/youtube_transcriber.py: Downloads and transcribes video audio via Whisper API
• tools/summarization.py: Summarizes the transcript using DeepSeek
• tools/web_search.py: Searches DuckDuckGo for current context
• tools/journalistic_highlight.py: Generates public interest highlights
• tools/index_transcript.py: Splits transcript and indexes it with FAISS
• tools/rag_query.py: Performs RAG query and allows fallback to general LLM knowledge
• tools/__init__.py: Makes the tools importable as a module

🔹 Requirements

• requirements.txt: All Python dependencies (tested with Python 3.12)
• packages.txt: System dependencies (e.g., ffmpeg, build tools)

---

📈 Improvements & recommendations

Here are potential enhancements for the project:

✅ Already solved

• Fixed repeated video download when switching tabs
• Added session state to persist transcript/vectorstore between tabs
• Enabled mixed-source RAG answers (video + general knowledge)
• Adjusted requirements.txt and packages.txt for compatibility

🚧 Future improvements

1. Caching

   • Save FAISS vectorstore, summary, highlights to disk (.save_local()) or use st.cache_data()

2. Better Streamlit UX

   • Enable chat-style Q&A with memory
   • Show progress for each processing step
   • Add "Download PDF" report button

3. Model prompting

   • Add clear tags like [FACT FROM VIDEO] vs [LLM KNOWLEDGE]

4. Testing

   • Add unit/integration tests using pytest

5. Agent orchestration

   • Split into two agents: VideoAnalysisAgent and QAAgent
   • Optionally adopt CrewAI or LangGraph for more complex flows

---

📄 License

MIT License. See LICENSE file.

---

🙌 Credits

Developed by Reinaldo Chaves (@reichaves) — journalist, data scientist, and investigative technologist.

---

💬 Questions or feedback?

Open an issue or contact via GitHub.