Planning, Memory, and Retrieval
Free NVIDIA-Certified Professional: Agentic AI practice — 6 questions on Planning, Memory, and Retrieval, with explanations. No sign-up. Full 12-question mixed test →
An agent executes a 40-step autonomous research task spanning multiple tool calls and sub-agent handoffs. After step 25, the agent begins pursuing a tangential sub-goal that technically relates to an earlier tool output but no longer serves the original user objective. Which architectural pattern BEST prevents this goal drift in long-horizon planning?
Long-horizon agents drift because intermediate context dominates over the original objective. Hierarchical planning with periodic re-grounding checkpoints against a persistent, explicitly stored goal representation is the standard mitigation tested on the exam for maintaining goal fidelity across many steps.
A retrieval pipeline combines BM25 keyword search results with dense vector similarity search results to answer agent queries over a mixed technical/legal corpus. The two systems produce scores on incompatible scales. Which fusion method should be used to merge the two ranked lists without requiring score normalization?
RRF fuses ranked lists using only rank position (1/(k+rank)), sidestepping the need to normalize incompatible scoring scales between sparse and dense retrievers. This is the standard hybrid retrieval fusion technique tested for agentic RAG pipelines.
You are building the retrieval component for an agent that answers questions over long legal contracts where clauses frequently reference terms defined many paragraphs earlier. Naive fixed-size chunking with no overlap is producing poor grounding because definitions get separated from the clauses that use them. What is the BEST chunking configuration to fix this?
Section-aware chunking respects document structure (e.g., clause/definition boundaries), and overlap ensures cross-references aren't split apart mid-context. This directly addresses the failure mode of definitions being separated from usage, which is a core exam topic on chunking strategy selection.
An agent's long-term memory store contains two types of entries: (1) records of specific past user interactions with timestamps and situational context, and (2) generalized distilled facts learned across many interactions with no reference to when or how they were learned. Which memory types do these two categories represent, respectively?
Episodic memory stores specific, time-and-context-bound events (what happened, when), while semantic memory stores generalized, decontextualized facts and knowledge distilled over time — a core distinction tested for agentic memory architecture design.
An agent performing an extended multi-turn tool-use loop is approaching its context window limit, driving up both latency and per-call cost as the conversation grows. The agent still needs to recall key facts and decisions from early in the session to complete the task correctly. Which strategy BEST balances context-window constraints with retention of critical information?
Progressive/rolling summarization (compaction) preserves salient early information in condensed form while keeping recent context verbatim for fidelity, directly reducing token count and cost/latency without losing critical earlier facts — the standard exam-tested context management technique for long sessions.
A retrieval pipeline uses a bi-encoder to fetch the top 50 candidate passages from a vector store for each agent query, but downstream answer accuracy is still poor because the top few results are often not the most relevant ones. Which addition to the pipeline BEST improves precision at low added cost, given that cross-encoders are too slow to run over the full corpus?
Cross-encoders jointly encode query and passage for higher precision but are too expensive to run over an entire corpus. The standard pattern is bi-encoder retrieval for recall over the full corpus, followed by cross-encoder re-ranking over a small candidate set (e.g., top 50) to boost precision at low marginal cost — a core hybrid/re-ranked retrieval concept tested on the exam.
Ready for the real thing?
The full course: two full-length practice tests, video lessons for every exam domain, hands-on labs and detailed explanations.
$129.99 $34.99 with code FREETEST33 — valid through September 7.
Get my $34.99 deal →