Workflow Overview

This is a product description generation system based on the RAG (Retrieval-Augmented Generation) architecture. It receives product information requests via webhook, retrieves relevant context from a vector database, leverages AI to generate high-quality product descriptions, and logs the results into Google Sheets.

Workflow Architecture

Trigger Mechanism

• Webhook Trigger: POST endpoint /product-description-generator, receiving external requests for product description generation.

Data Processing Pipeline

1. Text Preprocessing

• Text Splitter: Splits input text into chunks.
  • Chunk size: 400 characters
  • Overlap: 40 characters
  • Ensures textual coherence and preserves context

2. Vectorization and Storage

• Embeddings (Cohere): Generates text embedding vectors using the Cohere embed-english-v3.0 model.
• Pinecone Insert: Stores vectorized text in the Pinecone vector database.
  • Mode: insert
  • Index name: product_description_generator

3. Retrieval and Generation

Vector Retrieval

• Pinecone Query: Retrieves relevant product information from the vector database.
• Vector Tool: Wraps retrieval capabilities as a tool for the RAG Agent.

AI Generation

• Chat Model (Anthropic): Uses Claude as the language model.
• Window Memory: Maintains conversation context to support multi-turn interactions.
• RAG Agent: Core agent node
  • System prompt: "You are an assistant for Product Description Generator"
  • Integrates vector retrieval results and conversation memory
  • Generates compliant product descriptions

Output and Monitoring

Successful Processing

• Append Sheet: Appends processing results to Google Sheets.
  • Document: Product Description Generator
  • Worksheet: Log
  • Logged field: Status

Error Handling

• Slack Alert: Sends error notifications to the #alerts channel.
  • Message format: "Product Description Generator error: {error message}"

Technology Stack

AI/ML Components

• Vector Database: Pinecone
• Embedding Model: Cohere embed-english-v3.0
• Language Model: Anthropic Claude
• Memory System: Window Buffer Memory

Integrated Services

• Webhook: HTTP POST endpoint
• Google Sheets: Data logging
• Slack: Alert notifications

Data Flow

Webhook Request
  ↓
Text Chunking
  ↓
Vector Embedding (Cohere)
  ↓
[Parallel Paths]
  ├─→ Vector Storage (Pinecone Insert)
  └─→ Vector Retrieval (Pinecone Query)
        ↓
      Vector Tool
        ↓
    RAG Agent (integrating retrieval results + conversation memory + Claude)
        ↓
    [Conditional Branch]
      ├─→ Success: Log to Google Sheets
      └─→ Failure: Slack Alert Notification

Key Features

1. RAG Architecture: Combines vector retrieval with generative AI to ensure description accuracy.
2. Context Preservation: Window Memory enables conversational interaction.
3. Chunked Processing: Intelligent text splitting handles long-form inputs.
4. Logging & Tracking: Google Sheets records all processing statuses.
5. Error Monitoring: Real-time Slack alerts enable rapid issue response.

Use Cases

• Automated e-commerce product description generation
• Bulk SKU description optimization
• Multilingual product copy creation
• Description rewriting based on historical data

Configuration Requirements

API Credentials

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

Pinecone Setup

• Index name: product_description_generator
• Index must be pre-created with dimensions matching the Cohere model