Multi-agent RAG and knowledge integration – Multi-agent systems (AI-500, lesson 12)

What you learn in this lesson

decide which agent retrieves which knowledge, and justify the choice between central retrieval, retrieval per specialist agent and knowledge base as a decision note

cut supplier contracts by sections and justify the size and overlap of the chunks

enforce tenant separation in the index with a filter that the code builds from the run

take results of an MCP server into the source ledger as a knowledge source of their own, with origin

measure retrieval accuracy with a test set before sizes and thresholds are changed

The group-travel team of Nordlicht Reisen GmbH relies on two bodies of knowledge: the 120 documents from AI-103 (travel terms) and 40 supplier contracts (M-06, assumption). recht has to know terms, unterkunft and transport need contract clauses of the hotels and ferries, the pruefer reads the draft against the same sources. With one agent, RAG is a search. With seven, the question arises who retrieves what, what each may see and who owns a source.

Prior knowledge. "AI track, lesson T4.3" cuts texts at headings, "T5.1 to T5.4" grounds answers, names sources, declines without a source and shows hybrid search, "T6.2" fences foreign text. "AI-103, lesson 2.2" and "5.2" show index, semantic and hybrid query; "AI-103, lesson 5.5" introduces the source ledger per question. That is not repeated. New: several retrievers, one index with several tenants, MCP sources and measuring.

New in this lesson. The contracts live in a second index lieferantenvertraege; reisebedingungen (AI-103, R-08) stays unchanged. Both indexes carry a field mandantId that is set only where a document contains data of a customer (M-06), otherwise null. Retrieval uses a function tool wissen_suchen(frage); which area an agent reads is determined by the code, not by the model: recht reads the terms, unterkunft and transport the contracts, the planer both. The source ledger from AI-103, lesson 5.5 applies per run and remembers which agent fetched which source and whether it came from the index or from an MCP server.

The requirement

All numbers are assumptions of the task.

No.	Requirement	Type	Measure	Weight	Source
2.5-01	Each agent reads only the areas of its role.	R	0 accesses outside the table	Must	F-04
2.5-02	No hit carries the `mandantId` of another customer.	R	0 foreign hits	Must	M-07
2.5-03	The `pruefer` sees the sources that the draft and the specialist agents relied on.	F	source ledger per run	Must	M-03
2.5-04	Answers of an MCP server count as data, not as an instruction.	R	fence, origin noted	Must	M-12
2.5-05	The right clause is among the first five hits.	Q	at least 90 % of the test cases; assumption	Should	F-06
2.5-06	Retrieval is testable without Azure.	Q	stub instead of search service	Should	F-06

This lesson belongs to course AI-500

You will see the rest of this lesson and its video once you have bought the course. The goals above show what you learn here.

€129.00

incl. VAT, one-time payment

Access to all of the course's lessons with their exercises, solutions and videos on this site, for 24 months from the purchase. A self-study course: no tutoring, no exams or grading by us, and no certificate.

Withdraw from contract here