Case study
Multilingual, multi-tenant AI chatbots: Rasa Standalone instead of mainstream models
schedule3 min
calendar_today11 December 2024
syncCaricamento...
In recent months, integrating artificial intelligence into business systems has become a necessity for improving efficiency and the user experience. However, many AI models such as ChatGPT, Google's Bard and Anthropic's Claude can involve high costs and a dependency on cloud vendors. In this context, Rasa Standalone is a valid, free and highly customizable alternative, well suited to building multilingual, multi-tenant chatbots without compromising on quality.
Why choose Rasa?
Rasa is an open-source platform for building advanced chatbots, which offers:
- Full control: With Rasa Standalone, all data and models stay within the company's infrastructure, ensuring privacy and security.
- Full customization: The platform lets you adapt responses, dialogue flows and natural language processing to the specific needs of each tenant.
- Lower costs: Being open-source, Rasa removes the costs of API calls and of licenses for proprietary models such as OpenAI's.
Multilingual integration
One of the most demanding aspects of chatbots is handling multiple languages. With Rasa, we implemented a system that can:
- Detect the user's language dynamically: Through natural language processing pipelines that include components such as
LanguageDetector. - Handle language-specific responses: Each language is handled through dedicated intents and stories, avoiding overlaps or ambiguity.
- Extend easily to other languages: Rasa's flexibility makes it possible to add support for new languages without rewriting the whole system.
Multi-tenancy with a middleware
Managing multiple tenants was one of the main challenges we faced. Thanks to a custom middleware, we were able to:
- Isolate tenant data: Each tenant has its own set of intents, entities and responses.
- Configure models dynamically: The middleware identifies the tenant from the incoming request and loads the appropriate model.
- Support scalability and performance: Thanks to Docker and orchestrators such as Kubernetes, it is possible to handle dozens of tenants without compromising response speed.
Comparison with mainstream models
- ChatGPT (OpenAI): Offers fluent, contextual answers, but is based on paid APIs and does not guarantee full control over the data.
- Bard (Google): Ideal for research tasks, but not well suited to complex customizations for specific cases.
- Claude (Anthropic): With its focus on safety, it is interesting but still limited in terms of customization compared with Rasa.
Advantages of Rasa Standalone
- Zero costs: No recurring costs for APIs or licenses.
- Unlimited customization: Every aspect of the chatbot can be adapted.
- Local integration: All models run on-premise, ensuring greater security and regulatory compliance.
Challenges and solutions
- Complex initial setup: We got past the initial learning curve with detailed documentation and automation scripts for deployment.
- Optimizing for multiple languages: Building pipelines for multiple languages required a careful balance between performance and accuracy.
- Model maintenance: Thanks to a CI/CD system, we were able to update models and configurations without interruptions.
Conclusion
Rasa Standalone proved to be a powerful and flexible solution for building multilingual, multi-tenant chatbots, avoiding the costs and limitations of mainstream models. This experience underlines the importance of adopting open-source tools to keep full control and make the most of company resources. For anyone looking for an affordable, customizable alternative, Rasa is a strong choice.