Our Consultants' Former Clients

Program
10 modules
Duration
2 to 5 days, by scope
Format
On-site, online or hybrid
Certificate
Attendance certificate; Boğaziçi University-endorsed on request

Where does a RAG system break?

The demo works; production is harder. The training walks through the six stages of a RAG system one at a time and, at each stage, shows how to measure, find and fix the most common mistake.

  1. Ingestion

    Bringing in PDFs, wiki pages and project documents, parsing and cleaning them.

    Common mistakeTables and headings are lost during parsing.

  2. Chunking and metadata

    Splitting documents into meaningful chunks; adding fields such as source, date and department.

    Common mistakeA table is split across two chunks, so the answer is incomplete.

  3. Embeddings and vector DB

    Choosing the embedding model and vector database by data volume and language.

    Common mistakeChoosing an embedding model that is weak on Turkish queries.

  4. Retrieval and reranking

    Hybrid retrieval that combines BM25 and semantic search; ordering the results with a reranker.

    Common mistakeThe right chunk is found but never reaches the top ranks.

  5. Generation

    Generating the answer from the retrieved context, tying every claim to its source.

    Common mistakeThe model adds information that is not in the sources.

  6. Evaluation and guardrails

    Measuring retrieval and answer quality; protecting against prompt injection and data leakage.

    Common mistakeQuality is never measured; users find the problems first.

RAG, fine-tuning or long context?

Not every corporate knowledge problem is solved with RAG. The training teaches how to choose between the three approaches based on the shape of the problem.

Need RAG Fine-tuning Long context
Information changes often Good fitThe index is updated; the model stays the same. Poor fitEvery change requires retraining. Partial fitAlways current, but every document is sent with every request; cost and latency go up.
Answers must cite sources Good fitEvery answer is tied to the retrieved chunks. Poor fitKnowledge is baked into the model; sources cannot be shown. Partial fitPossible, but the number of documents is limited.
Large document collection Good fitScales to millions of chunks. Poor fitDoes not reliably teach facts; it is for behavior, not knowledge. Poor fitExceeds the context window.
Style and format must be taught Partial fitTo a limited extent, through the prompt. Good fitThe best method for this job. Partial fitTo a limited extent, through examples.
Per-user access rights Good fitA document-level permission filter is applied during retrieval. Poor fitWhatever the model learns is open to every user. Partial fitPermitted documents can be picked per request, but scale is limited.

The three are not mutually exclusive: RAG can run on a model fine-tuned for style and format, with prompt caching for repeated context.

Who is it for?

For technical teams building enterprise search, document search or knowledge assistants. Working knowledge of Python is enough; LLM and RAG concepts are covered from the start.

  • Backend and software developers

    Teams that will build enterprise search, a document assistant or a knowledge service.

  • AI and ML engineers

    Engineers who want to take an existing RAG prototype to production with measurable quality.

  • Data scientists

    Data scientists who want to measure retrieval quality, design evaluations and improve results.

  • Technical team leads

    Leads who will make RAG architecture, cost and security decisions and guide their team.

Program

Ten hands-on modules, from the basic pipeline to agentic RAG. The weight of each module is set by your team's level and goals.

  1. Introduction to RAG and enterprise use

    Why RAG is needed and when it is preferred over fine-tuning; grounded answers with citations; the need for accuracy, freshness and traceability.

  2. Data preparation and retrieval basics

    Ingesting PDF, wiki and markdown, parsing and cleaning; chunking strategies, metadata design; choosing an embedding model and a vector database.

  3. The basic RAG pipeline

    The end-to-end flow from ingestion to answer; connecting retrieval results to the prompt, citing sources, and the role of retrieval in reducing hallucination.

  4. Hands-on mini RAG service

    A question-answering service with FastAPI, an LLM API and a vector database: document upload, embeddings, retrieval and generation.

    Workshop Every participant stands up their own mini RAG service.

  5. Prompting, evaluation and debugging

    Prompt structures and structured output for RAG; measuring retrieval with recall@k and MRR and answer quality with faithfulness and LLM-as-judge; golden sets and regression tests.

  6. Advanced RAG architectures

    Hybrid retrieval, reranking, multi-query and query rewriting; BM25 analyzers and stemming for Turkish; context compression and filtering; an introduction to GraphRAG.

  7. Agentic RAG

    How agents differ from classic RAG; tool calling, planning and multi-step knowledge access across several data sources.

  8. Workflow integration

    External API and webhook integration, rate limits and error handling; human-in-the-loop approval at critical steps.

  9. Production, security and guardrails

    API-first services, latency and token cost; a guardrail layer against prompt injection and data leakage; document-level permission filtering during retrieval and role-based access control.

  10. Capstone: an enterprise knowledge assistant

    Designing, building and presenting a knowledge assistant that answers with sources from documents across different systems.

    Output A working prototype that can be adapted to your organization's scenario.

Get a quote

Format and certificate

Certificate

Every participant receives a RuneLab Academy certificate of attendance. On request, the training is run together with Boğaziçi University Lifelong Learning Center and the certificate is issued with BÜYEM endorsement.

Boğaziçi Üniversitesi Yaşamboyu Eğitim Merkezi logosu

Frequently Asked Questions

What is RAG?

RAG, short for Retrieval-Augmented Generation, means a language model first retrieves the relevant information from the organization's own documents and then bases its answer on those sources. The model is not retrained; as documents are updated, answers follow, and every answer can show its source.

What do participants need to know?

Working knowledge of Python and knowing how to call an API are enough. LLM and RAG concepts are covered from the start in the first module; for experienced teams the program is weighted towards the advanced RAG and agentic RAG modules.

Why does the length vary from 2 to 5 days?

The program is shaped around the team's level and goals. A program focused on the basic pipeline and the mini RAG service takes two days; the full program with advanced RAG, agentic RAG, integration and the capstone project takes five.

Which tools are used?

Exercises use LangChain or LlamaIndex, embedding models such as BGE-M3, and Qdrant, Milvus, Weaviate, pgvector or FAISS; the options are compared for your use case. Evaluation uses RAGAS. If your organization cannot use cloud models, the exercises can also be done with open source models running on-premise.

Is our company data used in the exercises?

No. Exercises use sample document sets we prepare; no real company data is entered into any tool. The capstone project is designed around your organization's scenario.

How is it different from the Local LLM Deployment Training?

This training focuses on designing systems that answer from corporate knowledge with cited sources, and it applies to any model, in the cloud or on-premise. The Local LLM Deployment Training covers running the model itself on your organization's servers. Organizations whose data cannot leave their network can plan both programs together. Explore the Local LLM Deployment Training.

Why RAG training if we can use RuneDesk?

RuneDesk is a complete software product built by RuneLab, installed in your organization and ready to use right away. It comes with sign-in and permissions, an admin panel and source-cited answers; it connects to sources such as SharePoint, Confluence and SAP, to databases, and to channels such as Microsoft Teams or Slack; and it runs in the cloud or on-premises. The RAG training, by contrast, helps your team understand how such systems work, add RAG-based features to their own products and assess a RAG solution against the right criteria. Its aim is not to replace a ready-made product but to help your team make sound decisions in this area and build solutions for its own needs. RuneDesk for an enterprise document assistant, the training for your team's skills: the two paths complement each other.

Do you issue a certificate?

Every participant receives a RuneLab Academy certificate of attendance. On request, the training can be run together with Boğaziçi University Lifelong Learning Center and the certificate issued with BÜYEM endorsement.

Who are your trainers?

The training is delivered by experienced senior AI engineers and data scientists who run enterprise AI projects at RuneLab.

Is the content adapted to our organization?

Yes. In a short call before the training we learn about your team's level, your document sources and the scenarios you want to solve, and prepare the modules and the capstone project accordingly.

How is the price set?

We prepare a quote for your organization based on the number of participants, the format and how much the content is tailored. Fill in the form and we will get back to you within one business day.

Get a quote for your team

Tell us the number of participants, your team's technical profile, and your preferred format and dates; we will get back to you within one business day.

Get a quote