RAG System for Intelligent Knowledge

Unified intelligent tool through which any employee (engineer, operator, manager) could pose a question in natural language.

OpenAI Embeddings

FAISS

Python

Introduction

Kazakhmys Corporation, a large mining and metallurgical holding company in Kazakhstan, faced the challenge that employees across divisions could not quickly and reliably get answers from their extensive internal knowledge sources—regulations, protocols, reports, internal databases, instructions. The client wanted to build a unified intelligent tool through which any employee (engineer, operator, manager) could pose a question in natural language and receive an accurate, sourced answer drawn exclusively from up-to-date internal knowledge. The project involved creating an MVP prototype, with target for future piloting and further deployment and scaling.

Business Challenges

1

2

3

Heterogeneous and fragmented data formats

Ensuring response accuracy and trustworthiness

Latency, scalability, and concurrency

Architecture & Concept We designed a RAG approach (Retrieval + Augmented Generation) comprising:

• Retriever / Search component — finds relevant knowledge fragments from the internalknowledge base (semantic/vector search).
• Generator / Summarizer — given the user query plus retrieved fragments, generates a relevant answer in the form of query-oriented extractive multidocument summary.
• Filtering / Verification module — ensures the generated answer stays within permittedinternal sources and rejects speculative content.

Architecture & Concept

1. Data consolidation & preprocessing o Audit and inventory knowledge sources (what documents, databases, systems)
o Normalize formats, chunk documents, attach metadata (tags, date, author, division)
o Deduplicate, cleanse, and review for sensitive content

2. Embedding & indexing o Select embedding model (e.g. Sentence Transformers, OpenAI embeddings, or domain-adapted model)
o Index embeddings in a vector database (Milvus, Weaviate, Pinecone, FAISS, etc.)
o Tweak search semantics and ranking algorithms

3. Quality control / validation
o Check thresholds: refuse answers when evidence is insufficient
o Fallback logic (e.g. “I’m not confident—please consult an expert”)
o Monitor metrics: accuracy, answer rejection rate, user feedback

4. Pilot, feedback loop, iteration & scaling o Launch pilot version in a single division o Collect user feedback, refine retrieval, improve prompts

1) A user types a question in the chat or portal

2) The backend converts the query to an embedding, performs vector search → retrieves top N fragments

3) These fragments + the original query are passed to the LLM, which generates a response

4) The answer is returned to the user, along with citations / source links

5) Admin dashboard tracks usage, flags weak answers, shows statistics

Results

Reduced load on subject-matter experts

Internal ad hoc questions to experts dropped by30–50 %

High user satisfaction

Percentage of accepted / accurate answers
(per user feedback)

Dramatic reduction in time to get internal answers

Employees now receive relevant information in seconds

Strengthened organizational knowledge culture and use of data

Other Cases

Increasing revenue with AI implementation: Dodo Pizza

Through in-depth research and analysis, we identify opportunities for growth, target audience insights, and the most

Increasing revenue with AI implementation: Dodo Pizza

Through in-depth research and analysis, we identify opportunities for growth, target audience insights, and the most

22 months

7 specialists

22 months

7 specialists

Would you like to discuss your project?

We will help you grow your business by creating and implementing a technology strategy tailored to your needs.