QSOFT implemented a YandexGPT-based RAG bot for Boiron without Python or orchestrators

QSOFT shared details about implementing a RAG bot based on Yandex Cloud Agent for pharmaceutical company Boiron. The assistant operates on the existing PHP and WordPress website, answers user questions around the clock, and takes on part of the load that previously went to manual support.

Why This Was Needed

Boiron in Russia has a large and complex catalog: 130 homeopathic single-ingredient preparations and 10 complex medicines. For users, the problem is not just the number of product cards, but the structure of knowledge around them. The portal contains descriptions of individual products, thematic collections, and more complex sections like "ENT Protocol." Because of this, a visitor's question cannot simply be matched to a single product by keyword: the system needs to understand the context and guide the user to the correct section of the knowledge base.

The support load was also significant. According to QSOFT, the website receives over 300 thousand requests per year — that's over 800 per day. With such a flow, some questions were not resolved immediately, some inquiries were lost, and some had to be handled by specialists manually. For medical topics, this is especially important: the user expects a quick answer, but the answer must be based on accurate data, not the model's free improvisation.

How the Solution Was Built

Instead of a separate Python service and complex orchestration loop, the team used Yandex Cloud Agent with the YandexGPT model and embedded the assistant into the existing PHP and WordPress stack. Essentially, this is a RAG scenario: the bot first searches for a relevant fragment in the knowledge base, and only then formulates an answer for the user. This approach reduces the risk of "hallucinations" and helps keep answers within the bounds of verified content, which for pharmaceuticals is more important than beautiful phrasing.

The project solved several practical tasks at once:

• launching without a complete website rebuild and abandoning the current stack
• searching a large medical knowledge base while accounting for connections between sections
• round-the-clock request processing without queuing for specialists
• reducing the number of lost queries and manual handling

The choice of architecture itself is particularly important. In many RAG projects, a layer of orchestrators, intermediate services, and custom pipelines quickly grows around the model. Here, the team took the opposite approach: first they preserved the working CMS and backend combination, and then added an agent layer on top of it. For business, this is simpler to maintain, cheaper to launch, and more straightforward for teams that already have a working website and have no desire to build a separate ML platform from scratch.

What the Implementation Provided

The main effect of such an assistant is not in "replacing an operator," but in closing typical scenarios on first contact. Users find the needed medication, section, or protocol faster, and specialists are involved where human expertise is truly needed. This changes the economics of support: less time is spent on navigating the portal and fewer inquiries get stuck between the website form and manual processing.

The implementation domain itself is noteworthy. In medicine and related fields, you cannot allow the model to answer too freely, because the cost of an error is higher than in typical e-commerce. Therefore, the Boiron case is interesting not as a demonstration of yet another chatbot, but as an example of careful RAG application in a sensitive industry. The team did not try to turn the assistant into a universal consultant, but limited it to a reliable knowledge base and the task of accurate search within it.

What This Means

QSOFT's case shows that applied AI for medicine can be deployed without a radical change in stack and without complex orchestration around the model. If a knowledge base is well-structured and search within it is carefully configured, even a PHP and WordPress website can gain a useful assistant that responds faster than people and stays within the bounds of verified content.