Optimise Shopping feeds with Generative AI. FeedGen is an open-source tool designed for merchants and advertisers to automatically improve product titles, generate comprehensive descriptions, and fill missing attributes in their Google Shopping feeds using Google's Large Language Models (LLMs). It streamlines the process of identifying and rectifying feed quality issues, aiming to enhance query matching, increase product coverage, and boost click-through rates.

How It Works

FeedGen operates as an Apps Script application within Google Sheets, providing a familiar interface for users. It leverages Google Cloud's Vertex AI API to access foundational LLMs, enabling both zero-shot and few-shot inference. The tool allows for customization through few-shot prompting, where users provide 3-10 examples from their own feed to tailor the model's output. It can also incorporate data fetched from product web pages and analyze product images using multimodal models for richer content generation.

Quick Start & Requirements

• Primary install / run command: Make a copy of the provided Google Sheets spreadsheet template and follow the instructions in the "Getting Started" worksheet. Authentication to the Apps Script environment is required.
• Non-default prerequisites and dependencies: Google Account, Google Sheets, Google Cloud Project with Vertex AI API enabled, and appropriate API authentication.
• Estimated setup time or resource footprint: Setup involves configuring the Google Sheet and Vertex AI access. The tool works best for up to 30,000 items; larger feeds may require alternative processing methods like BigQuery. Vertex AI API usage incurs costs.
• Links: Google Sheets spreadsheet template (via prompt), Getting Started worksheet (within template).

Highlighted Details

• Multimodal Image Understanding: Utilizes Gemini Pro Vision to analyze product images, generating higher-quality titles and descriptions by extracting visual features and cross-referencing with feed data.
• LLM Versatility: Supports various Google LLMs, including Gemini (1.5 Pro, 1.5 Flash, 1.0 Pro, 1.0 Pro Vision) and PaLM (text-bison-32k).
• Web Page Data Integration: Can fetch and parse content from product landing pages to enrich title and description generation.
• Scoring and Validation: Implements a scoring system for generated content to assess quality and facilitate manual review or bulk approval, including checks for Merchant Center compliance.
• Structured Data Output: Generates structured_title and structured_description attributes compatible with updated Merchant Center specifications.

Maintenance & Community

Community engagement is highlighted through spotlights on articles and case studies from contributors like Krisztián Korpa and Alex van de Pol, and features on Think with Google. Local development is supported via Node.js and npm, with instructions provided for building and deploying the application.

Licensing & Compatibility

The README describes FeedGen as an "open-source tool" but does not explicitly state a specific license type. Compatibility is primarily with Google Sheets and Google Merchant Center, requiring access to Google Cloud's Vertex AI.

Limitations & Caveats

FeedGen is explicitly noted as "not an official Google product." It is optimized for feeds up to 30,000 items, recommending alternative solutions for larger datasets. Manual review and verification of generated content are strongly advised due to the evolving nature of LLMs and potential for inaccuracies. Users must confirm that their target feed language is supported by Vertex AI.