Badou Unlocks Business Productivity by Building Enterprise Intelligent Chatbot Capabilities

By Dai Wei – Partner Solutions Architect, AWS
By Cui Xingnian – Co-founders, Badou Intelligent Technology Co., Ltd.

An enterprise knowledge base is a comprehensive, centralized repository of information designed specifically for use within a large organization. It serves as a single source of truth for customers and employees to access critical company knowledge, processes, and resources.

To improve the general search experience, traditionally, enterprises utilize a chatbot solution as the first-line support to extract information and a more human friendly user interface. It can effectively handle simple queries, provide basic customer support, and improve overall operational efficiency. But these chatbots often struggle with understanding context, nuance, or variations in phrasing, they rely on predefined Q&As and decision trees, which resulting in the construction of an enterprise knowledge base, requires a significant amount of manpower. They do not learn from interactions or improve their responses over time. These limitations highlight why many organizations are moving towards more advanced AI-powered chatbot solutions that offer greater flexibility, understanding, and learning capabilities.

Badou Intelligent Technology Co., Ltd. is a high-tech enterprise driven by AI technologies such as large language models, natural language processing, and knowledge graphs. With its comprehensive enterprise-level intelligent interaction products and applicable solutions, the company creates customized intelligent knowledge interaction platforms for enterprises. Badou has become a very important advanced tier partner of AWS in the field of generative AI, and is one of the earliest partners to obtain AWS Generative AI Software Competency, which demonstrated the highest level of generative AI specialization, deep AWS technical expertise, and proven customer success.

In this post, we will discover how Badou unleashing the business productivity with its innovative intelligent chatbot capabilities build on AWS.

Private Large Models Creation with Exclusive Instruction Data

Badou has built a large language model robot service platform, with its core being a 13B privatized exclusive large language model tailored for enterprises. This platform also allows free selection and switching of external third-party models.

Badou has implemented detailed supervised fine-tuning of the open-source model. The premise and core of this process is the construction of a large number of high-quality instructions sets to improve the basic model’s instruction alignment ability in specific scenarios.

The construction of fine-tuning instruction sets needs to follow these principles to achieve autonomous control of the large model:

1. • Diversity: Instructions should be diverse, allowing the model to learn the underlying logical paradigms of tasks.
   • Multi-domain: The instruction set should cover multiple industries and domains, giving the large model’s capabilities universality.
   • Professionalism: For specific domain-exclusive business capabilities, large model fine-tuning is necessary to achieve the expected results.
   • Consistency: The large model’s answer style should be consistent, allowing the model to have a unified approach to answering questions.

Based on the principles mentioned earlier, Badou has constructed a total of 1.14 million human-aligned instruction sets(Table 1). This extensive collection of instructions is designed to improve the large language model’s performance and alignment with human intent across enterprise domains and tasks.

Model fine-tuning based on open-source general models

Badou is specialized on model fine-tuning. The base model is trained on terabyte-scale linear text as a language model, learning knowledge in general domains, mainly focusing on concepts, entities, semantic and related knowledge, in order to achieve broad erudition. Badou utilizes a comprehensive set of training task instructions, through two technical approaches: RLHF (Reinforcement Learning from Human Feedback) and Self-Instruct, to derive specific capabilities from the language model. With only a small number of training samples, this enables the language model itself to possess integrated and coherent abilities, making the language model’s output consistent with human expectations. The fine-tuning process is illustrated in Figure 1.

Figure 1 – PreTrain And Human Alignment

Intelligent Q&A Integrated with Large Models

With its private large models, Badou then developed the intelligent chatbot capabilities by combining traditional chatbot functionalities with RAG framework.

Retrieval Augmented Generation (RAG) is an AI framework that enhances the capabilities of Large Language Models (LLMs) by incorporating external knowledge sources. RAG significantly enhances the capabilities of LLMs by providing them with access to current, relevant, and authoritative information, resulting in more accurate, trustworthy, and contextually appropriate responses while offering greater flexibility and control to organizations implementing AI solutions.

The following diagram(Figure 2) illustrates the workflow how Badou integrating large language models (LLMs) with local knowledge bases to create an intelligent chatbot system. Let’s break down the key components and processes depicted:

The flow starts with a user submitting a query to a Chatbot. The Chatbot first checks against a “Question and Answer Knowledge Base” to see if there’s a direct match for the query. At the first decision point, the system makes a judgment based on the knowledge matching score. If the score is high enough, it proceeds to “Accurate Answer”, providing a high-confidence response based on the local knowledge base. If not, it moves to the next step, employing a “Semantic Search Model” which utilizes both a “Question Knowledge Base” and a “Document Library” to find relevant information.

Another judgment is then made based on the search results. If satisfactory results are found, it moves to a “Private Large Model” for processing and generating a domain-specific answer. If there is no satisfactory results, it proceeds to another decision point, where a final judgment is made. This judgment is likely based on authorization or sensitivity of the query. If approved, it utilizes a “Public Large Language Model” to retrieve the most relevant result. If not approved, it falls back to the “Private Large Model”.

Depending on which path was taken, the system generates a response: “Accurate Answer” for high-confidence local knowledge matches, “General Response” based on the local knowledge base and the private large model, or “General Response” from the public language model (with a note that it’s for reference).

Figure 2 – Intelligent Chatbot Workflow

This workflow demonstrates an approach to chatbot responses, prioritizing local, verified knowledge when available, and falling back on more general LLM capabilities when necessary, all while maintaining control over information access and generation.

Benefits of the solution

By combining traditional NLP chatbot, RAG, private language models, and public LLMs, Badou created a powerful, flexible, and secure Enterprise Intelligent Chatbot solution(Figure 3) that significantly enhances both customer experience and operational efficiency.

Figure 3 – Badou Chatbot UI

This solution allows the chatbot to access up-to-date, company-specific information while leveraging both proprietary and broad general knowledge, resulting in more contextually relevant responses. The integration with enterprise knowledge base enhances information retrieval efficiency, enabling quick access to corporate documents and internal resources. Additionally, this system can be customized for specific industries or business functions, providing flexibility and adaptability to various organizational needs.

It also brings advantages in terms of data security, cost-effectiveness, and continuous improvement. By using private language models, sensitive company information remains protected, while RAG helps minimize incorrect information by grounding responses in verified data sources. This approach automates routine tasks, reducing the workload on human agents and decreasing the need for constant model retraining. The system leverages machine learning to improve responses over time and can be updated with new information without full model retraining. Furthermore, it integrates seamlessly with existing enterprise systems, providing data-driven insights from customer interactions and supporting business decision-making across departments, ultimately enhancing both employee services and customer experience.

Conclusion

Implementing an efficient Enterprise Intelligent Chatbot system can be complex, but Badou’s LLM Application Platform offers an out-of-the-box solution tailored for enterprise-level demands. It provides organizations with a cost-effective way to leverage large language models, enabling rapid deployment, customization, and scalability while ensuring data security and compliance.

Badou – AWS Partner Spotlight

Badou is an AWS Technology Partner and AWS Generative AI Competency Partner that supports internal business functions such as IT, HR, and Finance within enterprises. Based on departmental policy documents and other materials, and leveraging the semantic understanding capabilities of large language models, it builds an enterprise-level intelligent interaction platform to serve internal employees.

Contact Badou ｜AWS Marketplace