AI Inventory & Product Search Assistant
A deployed AI solution that transformed a leading hardware retailer's customer service, handling thousands of inquiries and streamlining sales through WhatsApp.
The Challenge
Operational Bottlenecks & The Solution
A leading hardware retailer faced significant operational bottlenecks on its primary digital sales channel: WhatsApp. Staff were overwhelmed by the high volume of customer inquiries coming through their WhatsApp channel.
This manual process led to slow response times during peak hours, inconsistent information, and took valuable time away from staff who could be assisting in-store customers.
Before
Human agents were tied up answering repetitive queries, manually checking inventory, calculating totals, and transcribing orders, leading to delays and potential errors.
The Solution
A deployed 24/7 AI assistant that instantly manages customer conversations, performs semantic product searches, and systematically captures all necessary data to formalize orders, integrating directly into the sales process.
System Architecture
Core Capabilities & Data Flow
Omnichannel Ingestion
The system captures customer inquiries in real-time via a WhatsApp Business API webhook, providing a frictionless entry point for users.
Intelligent Session Management
Using Redis, the architecture buffers and concatenates fragmented messages, ensuring the AI processes complete thoughts rather than isolated texts.
LLM Orchestration
A Gemini-powered agent acts as the central router. Governed by strict system prompts, it determines intent, maintains context, and decides when to invoke external tools.
Semantic Product Retrieval (RAG)
When product queries are detected, the agent queries a Qdrant vector database. This allows for natural language searches across the entire catalog, matching vague requests to precise SKUs.
Automated Order Fulfillment
Upon order confirmation, the agent executes a function call that structures the transaction data, logs it into a Baserow CRM, and alerts the fulfillment team.
Human-in-the-Loop Handoff
The system recognizes when a transaction requires human intervention (e.g., payment processing) and seamlessly transfers the context to a human advisor.
Under the Hood
Technical Deep Dive
01
Automated RAG Pipeline
The system's domain expertise is maintained via an automated Retrieval-Augmented Generation (RAG) pipeline. It continuously ingests the product catalog, generates high-dimensional vector embeddings (via OpenAI's text-embedding-3-large), and indexes them in a Qdrant vector database for sub-second semantic retrieval.
02
Human-in-the-Loop Escalation
The architecture features a deterministic escalation protocol. When the AI detects intent for payment or complex support, it triggers a sub-workflow that structures the conversational context, logs the pending transaction in the CRM, and alerts human agents via a dedicated channel, ensuring zero context loss during handoff.
03
Behavioral Prompt Architecture
The core LLM is governed by a sophisticated behavioral architecture defined within its system prompt. This enforces strict guardrails on tone, prevents hallucination of stock levels, and dictates a conversational cadence optimized for asynchronous messaging platforms like WhatsApp.
markdown
== TONE & STYLE ==
- Professional, empathetic, and helpful.
- Do not make final decisions or confirm stock or orders. Always defer to a human advisor once the customer is ready to proceed.
== RESPONSE FORMAT ==
- Always respond in short, separate messages. Mimic the conversational style of WhatsApp.
- Do not group multiple questions or sentences into a single message.The Impact
Quantifiable Results
85%Of Queries Automated
24/7Customer Availability
< 3 secAverage Response Time
The implemented solution revolutionized the retailer's customer service, providing instant, accurate support around the clock. The automation has measurably increased staff efficiency, allowing them to focus on high-value tasks like closing complex sales and in-store service.
The AI now serves as the primary point of contact, handling the majority of interactions and seamlessly escalating to a human expert with a complete conversational context and a pre-filled order in the CRM.
Gallery
n8n Workflow Architecture
Main AI Agent Workflow