Call-Center Assistant
We offer a ready-to-use, powerful RAG engine for finding answers to customer queries as quickly as possible.
1 day
quick start
100%
answer accuracy
This is a perfect option for rapid deployment. DKB acts as a fully encapsulated, ready-made solution, efficiently handling all indexing and information retrieval processes.
Business Objective
To quickly retrieve accurate answers to customer queries based on internal regulations, whilst eliminating AI "hallucinations".
Solution
At the heart of the solution is DKB — a knowledge base that enables interaction with corporate data in a chat format. The solution works as follows:
- Data pre-processing (Jupyter/Airflow + MLFlow): The developer configures a document preparation pipeline. Where needed, an LLM can be used at this stage to trace the text processing workflow prior to ingestion (LLM traces). As this is an open preparation process, its metrics and quality are logged via MLFlow for monitoring purposes.
- Knowledge base construction (DKB + Ollama + vLLM): Prepared documents are uploaded to DKB via API. To generate vector representations of text (embeddings), DKB automatically calls Ollama — dedicated exclusively to efficient embedding model hosting — and also makes requests to vLLM (LLM inference server).
- Response generation (vLLM): When an operator raises a query, DKB retrieves the relevant passage from the documentation and forwards it together with the query to vLLM — a high-performance inference server for large language models — which then generates the final response.
- Access control and billing (LiteLLM): All requests to vLLM are routed exclusively through LiteLLM. This allows administrators to configure usage limits for each operator or technical account, as well as monitor which department or application is consuming the most resources.
- User workspace: The operator interacts with the system via DKB's built-in, ready-to-use UI. It can additionally be integrated with the client's existing CRM and call-centre systems for greater context and detail.
Other cases
Would you like to read the full version of the case?
Fill out the form and we will contact you immediately.
Cookies Policy
We use cookies to operate this Site, understand how it is used, and support our marketing activities. Accept all, reject all, or set your own preferences.
Your trusted
partner
