Workflow Overview

This is an automated system for sending direct messages to new Twitter followers, integrating RAG (Retrieval-Augmented Generation) technology and a vector database. The workflow intelligently processes new follower data, uses AI to generate personalized direct message content, and logs operational results into Google Sheets.

Core Features

Intelligent Direct Message Automation

• When a new Twitter follower appears, the system triggers the workflow via a webhook
• Uses AI technology to generate personalized direct message content
• Automatically sends welcome messages to new followers

RAG Technology Integration

• Utilizes the Pinecone vector database to store and retrieve contextual information
• Converts text into vector representations using the Cohere embedding model
• Supports semantic search for more accurate contextual understanding

Workflow Architecture

1. Trigger Layer

Webhook Trigger

• Receives POST requests at the endpoint: auto-dm-new-twitter-followers
• Serves as the entry point for the entire workflow
• Receives relevant data about new followers

2. Data Processing Layer

Text Splitter

• Splits input text into smaller chunks
• Chunk size: 400 characters
• Overlap: 40 characters
• Ensures contextual coherence when splitting text

Embeddings

• Uses Cohere’s embed-english-v3.0 model
• Converts text into numerical vector representations
• Enables semantic similarity calculations

3. Vector Storage Layer

Pinecone Insert

• Stores embedded vectors in the Pinecone database
• Index name: auto-dm_new_twitter_followers
• Mode: insert

Pinecone Query

• Retrieves relevant context from the Pinecone database
• Queries using the same index
• Provides background information for AI generation

4. AI Processing Layer

Chat Model

• Uses OpenAI’s language model
• Responsible for generating intelligent response content

Vector Tool

• Name: Pinecone
• Description: Vector context
• Supplies vector query results to the AI agent

Window Memory

• Maintains conversation history
• Ensures contextual coherence
• Supports multi-turn conversations

RAG Agent

• System prompt: "You are an assistant for Auto-DM New Twitter Followers"
• Processing type: defined text processing
• Integrates vector tools and memory to generate final responses

5. Output Layer

Append Sheet

• Logs processing results to Google Sheets
• Document ID: SHEET_ID
• Worksheet name: Log
• Operation: append new row
• Logged field: Status

Slack Alert

• Error handling mechanism
• Sends to channel: #alerts
• Message format: Auto-DM New Twitter Followers error: {error message}

Data Flow

Webhook receive → Text split → Vector embedding → 
                      ↓
                 Pinecone storage
                      ↓
Webhook receive → Window memory → RAG agent ← Vector query ← Pinecone
                      ↓                ↓
                 Chat model      Vector tool
                      ↓
                 Success → Google Sheets log
                      ↓
                 Failure → Slack alert

Technical Highlights

High Intelligence

• Uses large language models to generate personalized content
• RAG technology ensures accurate and relevant replies
• Automatically learns and adapts to user preferences

Strong Scalability

• Vector database supports large-scale data storage
• Modular design facilitates feature expansion
• Supports custom prompts and parameters

Reliability Assurance

• Comprehensive error handling mechanism
• Real-time Slack alerts
• Google Sheets logging

Use Cases

1. Social Media Marketing: Automatically welcome new followers to boost engagement
2. Customer Relationship Management: Establish initial contact and collect user feedback
3. Brand Promotion: Deliver brand messaging and guide user behavior
4. Community Management: Batch-process new members with personalized welcomes

Configuration Requirements

API Credentials

• Cohere API (embedding service)
• Pinecone API (vector database)
• OpenAI API (language model)
• Google Sheets OAuth2 (data logging)
• Slack API (error notifications)

Total Nodes: 12 nodes

• 1 trigger node
• 4 data processing nodes
• 3 AI/ML nodes
• 2 vector storage nodes
• 2 output nodes