Why Regex Fails at Google Taxonomy: Building a 98% Accurate RAG Agent

The Problem: "Is a 'Hot Dog' a Dog?" 🌭 ## The Solution: Retrieval Augmented Generation (RAG) 🧠 ## 1. The Architecture ## 2. The "Smart Retry" Pattern 🔄 ## 3. The Stress Test 📉 ## Code Snippet (The Retry Logic) ## Conclusion In Google Merchant Center, categorization is everything. If you misclassify a product, your ads stop running. Most feed tools use keyword matching (Regex). This is why 15-20% of products in large catalogs often sit in "Disapproved" purgatory. I built CatMap AI to solve this using Vectors, not Keywords. Instead of rules, we convert the entire Google Product Taxonomy (5,500+ nodes) into a Vector Index using OpenAI's text-embedding-3-small. When a product comes in ("Pallash Casual Women's Kurti"), we don't look for the word "Kurti". We look for the mathematical concept of the product in vector space. Here is where it gets interesting. Standard Vector Search fails on cultural terms. To fix this, we implemented an Agentic Loop: We ran this system against 2,000 real-world edge cases. Regex is dead for categorization. Context-aware AI is the only way to handle the complexity of modern e-commerce catalogs. If you want to test the API, I'm opening a Free Beta for developers. Link to CatMap AI Follow me for more Engineering Deep Dives into AI Agents. Templates let you quickly answer FAQs or store snippets for re-use. Are you sure you want to hide this comment? It will become hidden in your post, but will still be visible via the comment's permalink. Hide child comments as well For further actions, you may consider blocking this person and/or reporting abuse CODE_BLOCK: // Simplified Logic if (result.status === "Uncategorized") { const synonyms = await expandQuery(product.name); // AI Call const newContext = await VectorStore.search(synonyms); return categorizeWithContext(product, newContext); } Enter fullscreen mode Exit fullscreen mode CODE_BLOCK: // Simplified Logic if (result.status === "Uncategorized") { const synonyms = await expandQuery(product.name); // AI Call const newContext = await VectorStore.search(synonyms); return categorizeWithContext(product, newContext); } CODE_BLOCK: // Simplified Logic if (result.status === "Uncategorized") { const synonyms = await expandQuery(product.name); // AI Call const newContext = await VectorStore.search(synonyms); return categorizeWithContext(product, newContext); } - Rule: If title contains "Dog" -> Category: Animals > Pets > Dogs - Input: "Hot Dog Costume" - Result: Animals > Pets > Dogs ❌ (Wrong!) - Input: Kurti - Vector Match: Generic Clothing (Confidence: Low) - Attempt 1: Standard Search. Result: Uncategorized. - Trigger: Agent detects failure. - Action: Agent calls an LLM (gpt-5-nano) to "expand" the query. Prompt: "What is a Kurti? Give me synonyms." Response: "Tunic, Blouse, Shirt". - Prompt: "What is a Kurti? Give me synonyms." - Response: "Tunic, Blouse, Shirt". - Attempt 2: Vector Search with "Tunic Blouse Shirt". - Result: Apparel > Clothing > Shirts & Tops. ✅ - Prompt: "What is a Kurti? Give me synonyms." - Response: "Tunic, Blouse, Shirt". - Coverage: 100% (Up from 85%). - Accuracy: 98.3%. - Time per Row: ~200ms.