12. The Q&A agent | Semantic Layer course

Section 1

Three ways to use what you just built

Next section

Section 2

Where the agent runs and what wraps it

Next section

Section 3

The nine steps, in order

Next section

Section 4

Tuesday, answered

Next section

Section 5

Daniel's application, and the decision the agent refuses to make

Next section

Section 6

The tools, the interfaces, and the other two agents

Next section

Section 7

Tracing ACE back to the components

Hands on

Wire the whole thing up

45 min

Run a local Q&A agent that classifies intent against the ontology, generates SPARQL, executes against Fuseki plus Ontop, runs the reasoner, retrieves a document citation, and returns the full answer plus evidence plus citations plus reasoning_steps JSON for the default question and for Daniel's application.

make lab-up            # once, starts Postgres + Fuseki + Ontop
make lab lab12-qa-agent

Full instructions, starter files and the solution are in labs/modules/lab12-qa-agent/README.md. The lab checks its own work, so you will know if it passed.

Check yourself

Quiz

7 questions, pass mark 70%. Answers are explained either way, so a wrong guess is not wasted.

1 Which two of the nine steps in the Q&A agent flow are always invoked? (pick two)

Select all that apply.

Question Intent Classification Symbolic Reasoning Response synthesis and citation generation GraphRAG document retrieval

2 Why does the intent classifier ground entities in the ontology instead of passing them downstream as strings?

Grounded IRIs are shorter than natural-language phrases, which reduces token cost An IRI carries a definition, a version, provenance and the axioms attached to the class, so downstream steps operate on a thing with a meaning instead of a word to guess about SPARQL cannot accept string literals in a query It lets the planner skip the VKG service entirely

3 The guidance says query-time symbolic reasoning is often unnecessary. Why, and when do you still need it?

Because the foundation model can perform the reasoning instead; you need the reasoner only when the model's confidence is low Because the Knowledge layer reasons continuously, so materialized knowledge is already validated and inferred; you need it when live federated data is combined with the graph, to check consistency, infer across the combination, or backward-chain from a goal Because reasoning is too slow for request-time; you need it only in nightly batch Because SHACL validation replaces it for most questions

4 Daniel Okafor's application APP-9001 has credit score 640 and DTI 0.41. What does the agent conclude, and why?

EligibleApplication, because the verified income document condition passes Automatically declined, because two of R4's three conditions fail HighRiskApplication and not EligibleApplication, so it routes to human review rather than auto-declining, because R5 requires review and the automatic-decline floors of score 560 and DTI 0.55 are not met Nothing, because R4 and R5 contradict each other and the ontology is inconsistent

5 The AWS guidance insists that every capability the internal Q&A agent uses must also be exposed to external agents, systems and users. What breaks if you skip that?

The agent runs slower, because internal calls are not cached Semantic capability gets siloed inside one agent implementation, so any other team needing it builds its own worse copy with its own private definitions MCP tool discovery stops working Guardrails can no longer be applied at the entry point

6 In the CRO's answer, the agent reports 1,203 defaulted loans but also mentions 1,847 and 4,102. Why report all three?

Because the reasoner could not determine which definition was correct Because all three are legitimate measures of different things, and naming which definition produced which number is what the semantic layer was built to make possible Because the confidence score was below 0.95, so alternatives are always listed Because the metric definition mtr:DefaultedLoanCount has three variants

7 What does the act claim in the inbound JWT record, and why does the guidance want it?

The action the user is requesting, so guardrails can block disallowed operations The agent identity, so the audit log records that the request was the user's, made through a specific version of a specific agent The activity timeout, after which the session context graph is discarded The account the query is billed to