AI Chatbot Development for a Retail Software Provider

Challenge

The client wanted to strengthen their product offering with AI-driven capabilities that would make everyday work simpler and more efficient for end users. However, reaching this goal required overcoming several obstacles that directly impacted business value.

Key challenges the client’s team faced included:

• Time-consuming reporting — employees had to rely on technical skills or developer-prepared queries to extract insights from company data.
• Inefficient workflows — routine questions about sales, stock, or suppliers took significant time to resolve manually.
• User adoption risks — without a modern, intuitive interface, the solution could struggle to attract new clients and stand out in a competitive market.
• Cost concerns — existing third-party chatbot tools were either too limited or too expensive to be a sustainable option.
• Evolving requirements — the client’s vision was still forming, and many specifications had to be clarified and refined during development.
• Handling unstructured and large documents — standard parsers failed on files over 10 pages, requiring custom development.
• Integrating structured and unstructured data — the chatbot needed to combine insights from synchronized database tables and uploaded documents.
• Performance bottlenecks — ensuring acceptable response times under heavy queries was a recurring challenge.
• Scalability for future clients — the architecture had to support eventual multi-tenant use without extensive rework.

With the desire to move beyond traditional reporting tools and deliver an intelligent, user-friendly experience to their customers, the client turned to us to bring the chatbot idea to life.

Solutions

To address the client’s requirements, our team designed and delivered a modular AI-powered chatbot that could be seamlessly integrated into their inventory and sales management platform. The solution was built with scalability, cost-efficiency, and ease of use in mind, combining Google Cloud services, a modern frontend, and intelligent data handling.

Modular architecture

The system was structured as a set of independent yet interconnected modules, which ensured flexibility, easy scaling, and simplified future enhancements:

• Data ingestion and parsing — a secure upload flow with short-lived links allowы users to import files safely. A custom parser was developed to handle large documents exceeding standard model limits, enabling the chatbot to process certificates, reports, and images efficiently.
• Structured data synchronization — dedicated modules synchronized sales, inventory, and supplier tables from the client’s database with BigQuery, ensuring the chatbot always worked with up-to-date records.
• Vectorization and retrieval — documents and table data were embedded and stored in BigQuery’s vector index. Similarity search (cosine and Euclidean distance) was used to retrieve the most relevant context for each query.
• LLM-powered query processing — Vertex AI with Gemini models (including Gemini 1.5 Flash for fast responses) served as the reasoning engine. Queries were reformulated against the retrieved context, ensuring accurate, domain-specific answers.
• Frontend interface — a web app built with Next.js provided a responsive, browser-based UI. AG Grid was integrated to give users Excel-like capabilities for working with large data tables, supporting import/export and advanced filtering.
• Security and access control — single sign-on enabled users to authenticate via their existing systems. Pre-signed URLs ensured secure file uploads, and role-based access was built into the architecture.
• Scalability and performance — the backend services were deployed on Cloud Run, enabling serverless autoscaling for handling query spikes. Performance improvements included caching frequently asked questions and categorizing documents to optimize search efficiency.

Implementation highlights

• Cost-efficient AI integration — by combining Vertex AI with in-house retrieval, the solution avoided the high per-query costs of third-party chatbots, making the system financially sustainable.
• Future-ready design — the architecture was built with multitenancy in mind, preparing the client to onboard new customers with minimal additional development.
• Developer-friendly environment — unit tests and modular design ensured quick iterations, while the use of widely adopted frameworks like Next.js and AG Grid simplified maintenance and onboarding.
• User-centric interface — the chatbot was embedded directly into the client’s product workflow, allowing users to move seamlessly from their system to the chatbot without switching tools.

By combining these modules, the chatbot delivered a smooth and reliable experience, enabling users to interact with both structured and unstructured data as naturally as asking a colleague.