AI Training · Enterprise RAG

Enterprise RAG Training for AI & Data Teams

Your team can wire a vector store to a prompt template and have a RAG demo running by lunch. Production retrieval-augmented generation lives or dies on chunking discipline, retrieval quality, grounded citations, and evaluation, the parts most teams skip until something breaks. This hands-on training takes AI and data engineers through the full pipeline: corpus design, hybrid retrieval and reranking, grounded generation, and the metrics that tell you whether a change actually helped. You leave with a reference architecture and an evaluation suite for your own corpus.

modules: 6
hours: 15

Most teams have already tried the fast version: wire a vector store to a prompt template and watch it answer five demo questions before calling the pilot proven. It rarely holds up once production traffic arrives. Nobody chunked the corpus on purpose or reranked results. They also built no way to tell whether an embedding-model change helped. Production retrieval-augmented generation lives or dies on the parts a demo skips: chunking discipline, retrieval quality, grounded citations, and evaluation.

This is a build session. Over two days, your AI and data engineers work the full pipeline on a slice of your own corpus. They design chunking and metadata for corpus documents before building hybrid search with reranking. Answers stay grounded through citations and multi-hop retrieval. Teams replace eyeballed test queries with a golden dataset that measures retrieval and generation quality.

A full module covers what production demands once the demo is over. That includes access control that travels with the document past ingestion and prompt-injection risk from retrieved content, plus live-system monitoring. We draw the boundary early. This training skips end-user research in a notebook tool and stand-alone prompt engineering. It also skips fine-tuning a model's weights. Those are different problems for a different session.

If your team is building agents that use retrieval among several tools, AI agents training is the natural next step. The AI training catalog shapes each program around your stack and data within your constraints. You leave with a reference architecture and an evaluation suite that keeps running against your own corpus after the trainers leave.

The demo takes an afternoon. Production takes discipline

Any team can wire a vector store to a prompt template in an afternoon. What holds up under real traffic is chunking discipline, retrieval quality, and evaluation, the parts most teams skip when the demo already looks convincing.

Retrieval decides everything

A perfect prompt can't rescue a wrong chunk. Most RAG failures trace back to retrieval: the wrong document surfaces, a chunk gets cut mid-sentence, nothing gets reranked, and the model just answers confidently with what it was given.

When the wrong policy version gets cited

A RAG system that cites an old policy version, or answers from data it should never have touched, is usually missing metadata, freshness handling, or an access-control rule at retrieval time.

Is anyone measuring retrieval before it ships?

Most teams eyeball a handful of test queries and call it done. There's no golden dataset behind that, and no retrieval or generation metrics either. So when someone changes the chunking or swaps the embedding model, nobody can say whether it helped.

Access control travels with the document

A retrieval system that ignores who's allowed to see a document turns your knowledge base into a leak waiting to happen. Permissions and tenancy must account for prompt-injection risk from retrieved content day one.

01
Where RAG helps, and where it doesn't
90 minBeginner
Retrieval-augmented generation earns its cost in some places and just adds latency in others, and this session draws that line before your team builds anything. We set one shared vocabulary for RAG versus fine-tuning versus a longer context window. That lets the rest of the course focus on implementation.
What RAG actually does, and where it quietly breaks
RAG vs. fine-tuning vs. long context
When a prompt alone is enough
Chatbot demo vs. production-ready RAG
Good first use cases: support and internal docs
02
Corpus design and the data pipeline
150 minIntermediate
This is the unglamorous work that decides everything downstream. You audit the corpus and pick a chunking strategy for each document type. Metadata then carries access rules and freshness with every chunk. The practice uses a slice of your own documents.
Auditing your corpus before you touch it
Chunking prose, tables, code, and structured docs
Cleaning and dedup supported by metadata for retrievability
Access and freshness tags at ingestion
PDFs, HTML, Confluence, and code repos
Your own ingestion pipeline
03
Retrieval engineering: hybrid search and reranking
180 minIntermediate
Naive vector similarity gets you a demo. This module builds an actual retrieval stack: choosing an embedding model, blending keyword and vector search, rewriting queries, and reranking before anything reaches the model. It's the core build session, and most of the training's lab time lives here.
Picking an embedding model and vector store
Hybrid search: BM25 plus vectors
Rewriting queries the way actual users ask
Reranking before generation
Metadata filters and tenant boundaries
Why the right document isn't in top-k
04
Grounding answers through citations and synthesis
150 minIntermediate
Retrieved passages become an answer the model can be trusted to give. Grounded prompts and citation formatting share context-window budgeting when several sources compete for space. We also cover multi-hop retrieval for questions one lookup can't answer, and what to do when retrieval comes up empty.
Grounding prompts in retrieved context
Citations that trace back to source
Budgeting context when sources compete
Conflicting or outdated sources
Multi-hop retrieval for harder questions
When retrieval comes up empty
05
Evaluation that isn't just eyeballing answers
150 minAdvanced
Eyeballed spot checks get replaced with defined measurement. Your team builds a golden evaluation dataset from representative questions. It includes retrieval metrics such as recall@k and nDCG, plus generation metrics for faithfulness. You leave with a regression check to run before every corpus or prompt change.
Recall@k and MRR alongside nDCG
Building a golden dataset from representative questions
Faithfulness and relevancy behind groundedness
RAGAS-style eval vs. human review
Regression testing before every change
06
Security and observability through production rollout
150 minAdvanced
Production trust starts with access control that travels with the document past ingestion, then adds protection against prompt injection from retrieved content, live monitoring, and a pilot-to-production plan with evaluation gates.
Enforcing access control at retrieval time as well as ingestion
Prompt injection from untrusted documents
Monitoring quality and latency live
Reindexing and versioning without downtime
Cost and latency tradeoffs
A rollout plan with evaluation gates

What you will learn

Tell RAG apart from fine-tuning and long context
Design a chunking scheme for corpus documents
A retrieval stack with hybrid search and reranking
Citations grounded in the source
Measure retrieval and generation quality with defined metrics
Catch access-control gaps before release
Plan a rollout with evaluation gates

Who should attend

AI/ML engineers building or maintaining retrieval systems
Whoever owns the ingestion pipeline on your data team
Solutions architects scoping a RAG approach
Applied AI teams shipping internal knowledge assistants
The tech lead setting your RAG standards
MLOps and DevOps, once the system reaches production

RAG teams train for two days, usually in half-day sessions that don't take a whole sprint off the calendar. The syllabus is the same onsite and live online. Bring a representative slice of your corpus for the build and evaluation modules.

Format: Onsite or live online
Duration: 2 days (about 14 hours, can be split into half-day sessions)
Group size: Up to 16 participants per group
Language: English or Turkish
Materials: Reference RAG architecture, evaluation harness templates, and a takeaway code repository
Certificate: Certificate of completion

Zeo started in 2011 and now works out of San Francisco, Istanbul, Ankara, and Lisbon. We run Copilot Academy and organize Digitalzone, an international digital marketing conference. This program draws on the 10+ years of consulting and training work behind that, applied to corporate AI adoption.