AI Meeting Summarizer

This repository contains the n8n workflows for a sophisticated AI Meeting Summarizer. This agent-based system allows users to process, summarize, and query Zoom and YouTube meetings through a simple chat interface.

The system is designed to be the central hub for all meeting-related information, leveraging AI to extract structured data and a dual-vector-database strategy for both broad and detailed semantic search.

Key Features

• Chat-based Interface: Users interact with a simple chat trigger.
• Multi-Platform Support: Processes both YouTube and Zoom meeting URLs.
• Intelligent Orchestration: A main AI agent analyzes user intent to decide which tool to use.
• Automated Transcription & Processing: Calls external services to scrape transcripts from provided URLs.
• AI-Powered Insights: Uses gpt-4o-mini to generate:
  • Executive Summaries
  • Key Action Items (with owners)
  • Structured Insights & Decisions
  • Attendee Lists
• Duplicate Prevention: Checks if a meeting URL has already been processed before starting a new transcription job.
• Dual Vector Storage: Creates two types of vector embeddings for nuanced search:
  1. Summary Embeddings: For broad, fuzzy search ("Which meeting was about Project X?").
  2. Transcript Chunk Embeddings: For detailed, granular Q&A ("What did Alex say about the Q4 deadline?").
• Conversation Memory: Remembers the context of the user's conversation using a Postgres database.

---

System Architecture

The project consists of two main n8n workflows that work together:

1. Chat agent.json (The Main Orchestrator): This is the user-facing workflow that manages the conversation and orchestrates all tasks.
2. Transcript_maker.json (The Processing Tool): This is a sub-workflow, run as a "tool" by the main agent, that handles the heavy lifting of transcription and database population.

---

1. Main Workflow: Chat agent.json

This workflow acts as the "brain" of the operation. It's a stateful agent that maintains a conversation and intelligently calls one of its three available tools based on the user's request.

Core Components:

• ** Chat Trigger:** Provides the public-facing chat interface for the user.
• ** AI Agent (Meeting Orchestrator):** The central agent (gpt-4o-mini) that interprets the user's needs. Its system prompt instructs it to follow a specific, step-wise logic.
• ** Postgres Chat Memory:** Remembers the previous messages in the conversation.

The Agent's Tools:

1. ** Meeting Existence Checker Tool (HTTP Request)**

   • Purpose: To prevent redundant processing.
   • Trigger: Called whenever a user provides a new URL.
   • Action: Performs a GET request to the Supabase meetinglist table to check if a record with the zoom_url already exists.
   • Logic:
     • If Found: The agent informs the user the meeting is already processed and proceeds to the Retrieval Tool.
     • If Not Found: The agent proceeds to the Transcript Maker Tool.

2. ** Transcript Maker Tool (Workflow Tool)**

   • Purpose: To process a new meeting.
   • Trigger: Called when a user provides a new, unprocessed URL.
   • Action: Executes the entire Transcript_maker.json sub-workflow, passing the URL as a parameter.

3. ** Meeting Retrieval Tool (Supabase Vector Store)**

   • Purpose: To answer questions about meetings.
   • Trigger: Called when a user asks a question without a URL (e.g., "What were my action items last week?") or when they ask a question about a meeting that is already known to exist.
   • Action: Performs a semantic search against the transcript_chunks vector database to find the most relevant information to answer the user's query directly.

2. Sub-Workflow: Transcript_maker.json

This workflow is the "factory." It is not triggered directly by the user but is called by the Chat agent. Its sole purpose is to take a URL, process it, and populate the databases.

Step-by-Step Flow:

1. Receive URL: The workflow starts with the URL provided by the main agent.
2. Classify URL: An AI model (gpt-4o-mini) analyzes the URL and outputs a JSON object identifying it as {"task": "YouTube", ...} or {"task": "Zoom", ...}.
3. Scrape Transcript: A Switch node routes the URL to the correct scraper:
   • Zoom: Calls a custom Google Cloud Run service (zoom-transcript-scraper...) to get the transcript.
   • YouTube: Calls a custom Google Cloud Run service (youtubecc...) to get the transcript.
4. Generate AI Insights: The raw transcript is fed to gpt-4o-mini with a specific prompt to extract structured JSON containing the summary, action items, insights, attendees, and metadata.
5. Populate Databases (Dual Storage Strategy): This is the most critical step.
   • Database 1: meetinglist (For Summaries)
     1. Insert Row: The AI-generated summary and metadata are saved to a primary table in Supabase (HTTP Request5).
     2. Create Summary Embedding: A new embedding is generated only from the AI-generated summary and title (LangChain Code1).
     3. Update Row: This "summary embedding" is added to the meeting's row in meetinglist (HTTP Request3). This embedding is used by the Meeting Existence Checker Tool for broad semantic matching.
   • Database 2: transcript_chunks (For Detailed Q&A)
     1. Chunk Transcript: The original, raw transcript is split into small, overlapping chunks (LangChain Code).
     2. Create Chunk Embeddings: An embedding is created for each individual chunk.
     3. Insert Chunks: Each chunk, its embedding, and the transcript_id (linking it to the main meetinglist row) are saved to the transcript_chunks table (HTTP Request1). This database is used by the Meeting Retrieval Tool for detailed Q&A.

---

Workflow Videos

• Workflow 01
• Workflow 02
• Workflow 03

---

Setup & Prerequisites

To run this project, you will need:

• n8n: An active n8n instance (cloud or self-hosted).
• OpenAI: An API key for gpt-4o-mini and text-embedding-3-small.
• PostgreSQL: A database for the Chat Memory node.
• Supabase Account:
  • A project with two tables:
    1. meetinglist: To store the main meeting record, metadata, and summary embedding.
    2. transcript_chunks: To store the raw transcript chunks and their individual embeddings.
  • A database function (e.g., match_meetings) for vector search, as referenced by the Meeting Retrieval Tool.
• Deployed Scraper Services:
  • This workflow relies on two external, custom-deployed services (likely on Google Cloud Run) to handle the actual transcript scraping:
    • https://youtubecc-423771082043.us-central1.run.app
    • https://zoom-transcript-scraper-423771082043.us-central1.run.app
  • You will need to deploy your own versions of these scraper functions for the workflow to be operational.