Extension Framework

AI DIAL presents a robust Extension Framework and plug-in infrastructure, enabling seamless integration of your data and business workflows with Language Models (LLM) to enrich your enterprise applications. Harness the full potential of our solutions to drive innovation and efficiency within your organization.

framework schema

Extension Framework Ecosystem


Addon is a service or a component, compatible with own Open API specification, that enables LLMs (Large Language Models) to utilize any desired data source or technology to produce their responses.

Conversational App

Applications are ready-to-use solutions, compatible with the DIAL API requirements, that combine configurations of Addons with other services, or any custom logic to achieve a specific system behavior.


Assistant is created by combining Addons and system prompts to achieve a specific behavior for the LLM, allowing for an enhanced flexibility and customization in its responses to meet specific requirements.

framework schema
EPAM offers custom development of AI DIAL extensions (Addons, Application, Assistants) to meet specific needs of your business leveraging extensive knowledge of AI-enabled custom data and workflows for GPT and other LLM models and with a strong emphasis on privacy and security, data science and data governance.

Key Factors for a Successful LLM Integration

Data Preparation

To correctly prepare data for LLM, it is vital to know how to handle specific data formats and have a deep insight into a business domain in order to assess what data has a higher priority. As LLM has a limited context capacity, it is crucial to input only carefully selected relevant data in the correct format, as this significantly impacts the quality of generated answers. You should have tools allowing to access and view the documents that are passed to the model to ensure they meet your goals.

Data Origins

To mitigate hallucination factors, it's essential to provide direct quotes and links to sources, allowing users to verify the information when it is needed. Such approach increases the user's trust and confidence that the AI-generated response is accurate and reliable.


To foster data security and privacy of a sensitive information, it is essential that a chat system conforms to an organization's existing data access policies to ensure that users can only access authorized data. Additionally, the chat system should be designed to handle access control when the data is changed or updated.


Implementing vector index management techniques is crucial to ensure the accurate and reliable retrieval of information. There is a potential risk that the a priori data utilized by the model to generate answers may be outdated, imprecise or incorrect. Therefore, it is essential to determine the proportion of a priori data versus verified data utilized by the model and prioritize the use of the most recent data.

Accuracy & Reliability

It is imperative to adopt a rigorous domain-specific quality assurance methodology to assess the accuracy and effectiveness of AI-generated answers, as this enables identification of potential issues or areas for improvement, ultimately enhancing the overall semantic performance of the AI system in support of your business objectives.

Precision & Hallucinations

It is essential to have the appropriate tools in place to evaluate the level of hallucinations (i.e., inaccurate or nonsensical responses) in AI-generated answers and to assess the level of precision in generating appropriate responses.

System Load Resistance

The system must be ready to perform under high load without experiencing significant performance issues or downtime and automatically adapt to an unexpected usage spikes. To achieve enterprise-grade scalability, the system architecture should incorporate advanced techniques such as load balancing and optimized resource allocation, while utilizing a blend of models to handle complex scenarios.


On an enterprise scale, the processing of large volumes of data may entail significant computational expenses. Additionally, some AI models may require a higher processing capacity than others, affecting the overall cost. Through meticulous analysis, businesses can leverage cost optimization strategies and empower AI that meets their need while still operating within the set budget.
When developing a Extension Framework, our primary emphasis is on the essential elements for a successful integration of LLM, considering your business requirements and the quality of responses.

Typical Addon Examples

Semantic Search
Query generation

Semantic Search Addon

Semantic Search

Golden Q&A Addon


Query Generation Addon

Query generation

Hybrid Solution for Conversational Search in the Context of Corporate Data