SummitAI

From Meetings to Meaning — AI that Transcribes, Summarizes, and Surfaces Insights.

SummitAI is an AI-powered meeting companion that transforms raw meeting recordings into structured knowledge.
Users upload a video file, and SummitAI automatically extracts the audio, transcribes speech using Whisper, and applies LLM-powered pipelines for summarization, question answering, and retrieval-augmented Q&A chat.

---

✨ Features

1. Video & Audio Processing

• Uploads MP4 video via Streamlit frontend.
• Load Video Segments using moviepy.
• Extracts audio (ffmpeg).
• Preprocesses audio (mono, 16kHz mono WAV) (Whisper-ready)
• Supports chunked transcription to avoid long input issues.

2. Transcription

• Speech-to-Text
  • Uses OpenAI Whisper API for accurate transcripts.
  • Returns structured transcript with:

    { "start": float, "end": float, "text": str }    

• Text Cleaning & Structuring
  • Removing filler words.

3. Meeting Summarization

• Implemented Map-Reduce Summarizations
  • map_prompt: chunk-wise summaries (main points + perspectives).
  • reduce_prompt: combines into Decisions, Action Items, Open Questions.
  • Output is concise, structured, and useful for meeting follow-ups.

4. RAG-based QA

• Chunk transcripts into vector DB. For developement, stored in InMemoryVectorStore.
• Embedding function is based on openAI's text-embedding-3-large.
• Question answering flow:
  • Retrieve relevant transcript chunks.
  • Pass context + question into LLM (via rlm/rag-prompt) - - - Return grounded answers.
• Natural language queries like: “What was discussed about budget in the meeting?”
• Only summary, transcript, and graph are stored in session (not raw video/audio).

4. RAG-based QA

• Upload & Process video (single click).

• Sidebar: Meeting Summary.

• Transcript Viewer: hidden under expander to avoid clutter.

• Chatbot Interface:

  • Persistent chat history via st. session_state.messages.
  • User can ask natural questions about the meeting.
  • Powered by RAG backend.

• Only summary, transcript, and graph are stored in session (not raw video/audio).

• Multi-Meeting Knowledge Base

  • Search across multiple past meetings.
  • Cross-meeting queries for recurring topics.

• Dashboard & Integrations

  • Web UI (Streamlit/Next.js).
  • Export summaries (PDF/Notion/Trello).
  • Slack/Email integration for automatic meeting notes delivery.

---

⚙️ Tech Stack

• Video/Audio Handling: moviepy, pydub
• Speech-to-Text: OpenAI Whisper
• LLM Framework: LangChain
• Vector Store: InMemoryVectorStore (easily changeable to Chroma)
• Embeddings: OpenAI text-embedding-3-large
• Summarization: LangChain Map-Reduce
• Frontend: Streamlit (dev)

---

📌 Project Status

• Phase 1 (Done): Video loading, audio extraction, test clip generation.
• Phase 2 (Done): Whisper transcription + cleaning pipeline.
• Phase 3 (Done): Summarization + RAG-based QA.
• Phase 4 (Done): UI dashboard + integrations.

---

📖 Example Use Case

• Upload a 30 min Zoom/Meet recording.
• SummitAI extracts and transcribes speech.
• Get a structured meeting summary with key decisions and action items.
• Ask questions like:

  "What did was discussed about the Q4 budget?"
  "List all action items assigned to Bob."

---

Special Notes

• If you want to change the clip length considered for the SummitAI, change the clip_length variable to desired time(in seconds) in extract_audio_from_video function of audio_extract.py