Advanced · 2 days
Agentic RAG Systems
A two-day advanced course combining retrieval with agentic reasoning and tool use: multi-step retrieval, query planning, structured sources and a production architecture for complex knowledge work.
2 days · Advanced · RAG & Agents · Experienced AI engineers, developers and architects
Overview
Static RAG answers a question with a single retrieve-then-generate pass. That is often enough. It fails when the question requires planning, multiple searches, comparison across sources, or a mix of documents and APIs. Agentic RAG is the attempt to do that work without turning the system into an unsupervised research assistant with no budget.
This course starts from the limits of static RAG, then builds the pieces of an agentic retrieval system: query planning, multi-step retrieval, tool-driven search, document selection, iterative retrieval and memory. Structured knowledge sources and APIs are combined with unstructured corpora. Planning, evaluation, security, latency and cost are treated as first-class design constraints.
The practical work is a complex system, not a toy. You implement a path that can plan, retrieve, call tools, keep intermediate state and produce a grounded answer with traces.
You leave able to design agentic retrieval where it is justified, and to keep static RAG where it is not.
Audience
Experienced AI engineers, developers and architects who already understand RAG and tool calling and need to compose them.
Prerequisites
Experience implementing RAG and tool calling, or completion of the RAG and agents courses. You should be comfortable with production application structure.
Duration
2 days
Best delivered against a customer corpus and internal APIs where those can be shared.
Learning outcomes
- 01
Diagnose when static RAG is the wrong architecture for a knowledge task.
- 02
Design query planning and multi-step retrieval without unbounded loops.
- 03
Combine unstructured retrieval with APIs and structured sources in one orchestration.
- 04
Implement iterative retrieval, document selection and memory of intermediate findings.
- 05
Evaluate agentic retrieval for completeness, groundedness, latency and cost.
- 06
Apply security controls so retrieved content and tools cannot expand authority.
- 07
Specify a production architecture with traces, budgets and fallback to simpler retrieval.
Outline
- 01
Where static RAG stops
- Multi-hop questions, comparisons, and tasks that need a plan before a search.
- The cost of stuffing more context instead of retrieving with intent.
- A reference architecture for agentic retrieval: planner, retrievers, tools, memory, generator.
- Exercise: take a failed static RAG question and write the retrieval plan a system would need.
- 02
Planning, tools and iterative retrieval
- Query planning, decomposition and choosing among search tools.
- Multi-step retrieval, document selection and stopping when evidence is sufficient.
- Tool-driven search: catalogues, APIs, SQL, and not treating every source as a vector index.
- Memory of intermediate findings without leaking them as unearned certainty.
- Lab: implement a planner that can retrieve more than once.
- 03
Mixed sources and grounded answers
- Combining unstructured documents with structured systems.
- Citations, conflict, and presenting uncertainty rather than averaging sources.
- Security: injection via retrieved text, tool results, and permissions on each source.
- Lab: add an API or structured lookup beside document retrieval.
- 04
Production architecture
- Latency and cost budgets for multi-step systems.
- Evaluation of plans, retrieval coverage and final groundedness.
- Orchestration, fallbacks to static RAG, and human handoff.
- Operational traces and a deployment shape.
- Build: complete a complex practical system and evaluate it against multi-hop questions.
Practical work
You build a multi-step retrieval system against a corpus that cannot be handled by a single search. The system plans queries, retrieves iteratively, can call at least one non-vector tool, and produces a grounded answer with traces and a cost/latency budget.
Takeaways
- A working agentic RAG system
- Planning and retrieval patterns for multi-hop questions
- Evaluation cases that static RAG fails and the agentic path should pass
- A production architecture with budgets, traces and fallbacks
Delivery
Cognitecta delivers private corporate training, on-site or as remote live training. Courses can be run as published, or adapted to your organisation’s stack, domain and experience level.
Instructor
Nicholas Johnson, AI architect and software engineer. He has a degree in Artificial Intelligence and around twenty years of professional technology training, including hundreds of courses for engineering teams and large organisations. About.
Related courses
2 days · Advanced
Building AI Agents
A two-day course for intermediate and advanced developers: design and build useful agents with tools, state, permissions and evaluation, rather than unsupervised demos.
2 days · Intermediate
RAG: From Fundamentals to Production
A two-day course on retrieval-augmented generation as an engineering discipline: why naive RAG fails, and how to design a system that can be evaluated, secured and operated.