Topic 03 Practice Test 3 covers Vector Stores, Metadata, Chunking, Embeddings, Retrieval, RAG, and Data Maintenance in Domain 1 (Tasks 1.4-1.5). For broader exam preparation, review the AIP-C01 Exam Dumps. Every option includes focused technical reasoning tied to the tested decision.
Question 1
An OpenSearch vector index is receiving high concurrent query volume and one partition has become the throughput bottleneck. The design must distribute vector-search work across the index. Which design choice best satisfies this requirement? Choose ONE.
- Monitor KNNGraphMemoryUsage and size OpenSearch data nodes for graph memory
- Use approximate k-NN with HNSW for the large vector index
- Route specialized domains to separate vector indexes
- Scale the OpenSearch vector index with an appropriate sharding strategy
Correct Answer: D
Correct Answer
Answer D is correct because This option is correct because it distributes index data and vector-query load across OpenSearch shards. That directly addresses one partition is the throughput bottleneck. Because one partition is the throughput bottleneck, this is the most direct design response.
Incorrect Answers
Answer A is incorrect because This option tracks OpenSearch k-NN graph memory against its circuit-breaker allowance. It fits when ANN graph memory pressure is causing vector-search instability. Here the decisive condition is one partition is the throughput bottleneck. In contrast, Scale the OpenSearch vector index with an appropriate sharding strategy addresses that condition directly.
Answer B is incorrect because This option uses approximate HNSW search to reduce nearest-neighbor latency at scale. It fits when exact vector search is too slow for the corpus size. Here the decisive condition is one partition is the throughput bottleneck. In contrast, Scale the OpenSearch vector index with an appropriate sharding strategy addresses that condition directly.
Answer C is incorrect because This option routes separate subject domains to independently tuned vector indexes. It fits when different corpora need independent tuning and lifecycle control. Here the decisive condition is one partition is the throughput bottleneck. In contrast, Scale the OpenSearch vector index with an appropriate sharding strategy addresses that condition directly.
Question 2
Legal, engineering, and product corpora have different update schedules and retrieval tuning. The platform wants to tune and operate each corpus independently while routing queries by domain. Which design choice best satisfies this requirement? Choose ONE.
- Store normalized domain-classification tags in metadata
- Scale the OpenSearch vector index with an appropriate sharding strategy
- Route specialized domains to separate vector indexes
- Use separate vector indexes for hard tenant isolation
Correct Answer: C
Correct Answer
Answer C is correct because This option is correct because it routes separate subject domains to independently tuned vector indexes. That directly addresses different corpora need independent tuning and lifecycle control. Because different corpora need independent tuning and lifecycle control, this is the most direct design response.
Incorrect Answers
Answer A is incorrect because This option normalizes business-domain labels across all ingested sources. It fits when source systems use inconsistent category names. Here the decisive condition is different corpora need independent tuning and lifecycle control. In contrast, Route specialized domains to separate vector indexes addresses that condition directly.
Answer B is incorrect because This option distributes index data and vector-query load across OpenSearch shards. It fits when one partition is the throughput bottleneck. Here the decisive condition is different corpora need independent tuning and lifecycle control. In contrast, Route specialized domains to separate vector indexes addresses that condition directly.
Answer D is incorrect because This option separates tenants into different vector indexes for storage-level isolation. It fits when policy forbids cross-tenant candidate visibility. Here the decisive condition is different corpora need independent tuning and lifecycle control. In contrast, Route specialized domains to separate vector indexes addresses that condition directly.
Question 3
A very large enterprise corpus needs a coarse-to-fine retrieval architecture that can narrow broad candidate regions before performing more detailed semantic matching. Which design choice best satisfies this requirement? Choose ONE.
- Use a hierarchical indexing architecture
- Use approximate k-NN with HNSW for the large vector index
- Route specialized domains to separate vector indexes
- Use fixed-size chunking with overlap
Correct Answer: A
Correct Answer
Answer A is correct because This option is correct because it narrows a large corpus through coarse and fine retrieval layers. That directly addresses a flat search layer no longer scales efficiently. Because a flat search layer no longer scales efficiently, this is the most direct design response.
Incorrect Answers
Answer B is incorrect because This option uses approximate HNSW search to reduce nearest-neighbor latency at scale. It fits when exact vector search is too slow for the corpus size. Here the decisive condition is a flat search layer no longer scales efficiently. In contrast, Use a hierarchical indexing architecture addresses that condition directly.
Answer C is incorrect because This option routes separate subject domains to independently tuned vector indexes. It fits when different corpora need independent tuning and lifecycle control. Here the decisive condition is a flat search layer no longer scales efficiently. In contrast, Use a hierarchical indexing architecture addresses that condition directly.
Answer D is incorrect because This option creates predictable token windows with overlap between adjacent chunks. It fits when regular documents favor deterministic segmentation. Here the decisive condition is a flat search layer no longer scales efficiently. In contrast, Use a hierarchical indexing architecture addresses that condition directly.
Question 4
Support answers must exclude obsolete procedures older than a policy-defined cutoff. Every ingested document has a reliable last-updated timestamp. Which design choice best satisfies this requirement? Choose ONE.
- Store authorship as filterable metadata
- Preserve source URI and citation metadata
- Store document freshness timestamps as filterable metadata
- Trigger the vector update pipeline from source-change events
Correct Answer: C
Correct Answer
Answer C is correct because This option is correct because it stores update time as filterable metadata for freshness rules. That directly addresses retrieval must exclude content older than a policy cutoff. Because retrieval must exclude content older than a policy cutoff, this is the most direct design response.
Incorrect Answers
Answer A is incorrect because This option stores author identity as a consistent retrieval attribute. It fits when queries and audits must constrain results by approved author. Here the decisive condition is retrieval must exclude content older than a policy cutoff. In contrast, Store document freshness timestamps as filterable metadata addresses that condition directly.
Answer B is incorrect because This option preserves provenance so retrieved chunks can point to their source. It fits when users need traceable evidence links. Here the decisive condition is retrieval must exclude content older than a policy cutoff. In contrast, Store document freshness timestamps as filterable metadata addresses that condition directly.
Answer D is incorrect because This option starts vector updates when source-change events occur. It fits when new procedures must become searchable within minutes. Here the decisive condition is retrieval must exclude content older than a policy cutoff. In contrast, Store document freshness timestamps as filterable metadata addresses that condition directly.
Question 5
Researchers often ask for guidance written by a specific approved team or author, and governance reviewers need the same attribute for audits. Which design choice best satisfies this requirement? Choose ONE.
- Store normalized domain-classification tags in metadata
- Store document freshness timestamps as filterable metadata
- Preserve source URI and citation metadata
- Store authorship as filterable metadata
Correct Answer: D
Correct Answer
Answer D is correct because This option is correct because it stores author identity as a consistent retrieval attribute. That directly addresses queries and audits must constrain results by approved author. Because queries and audits must constrain results by approved author, this is the most direct design response.
Incorrect Answers
Answer A is incorrect because This option normalizes business-domain labels across all ingested sources. It fits when source systems use inconsistent category names. Here the decisive condition is queries and audits must constrain results by approved author. In contrast, Store authorship as filterable metadata addresses that condition directly.
Answer B is incorrect because This option stores update time as filterable metadata for freshness rules. It fits when retrieval must exclude content older than a policy cutoff. Here the decisive condition is queries and audits must constrain results by approved author. In contrast, Store authorship as filterable metadata addresses that condition directly.
Answer C is incorrect because This option preserves provenance so retrieved chunks can point to their source. It fits when users need traceable evidence links. Here the decisive condition is queries and audits must constrain results by approved author. In contrast, Store authorship as filterable metadata addresses that condition directly.
Question 6
Documents arrive from several systems with inconsistent category names. Retrieval must apply one common business-domain taxonomy regardless of source. Which design choice best satisfies this requirement? Choose ONE.
- Route specialized domains to separate vector indexes
- Store authorship as filterable metadata
- Apply an explicit metadata filter
- Store normalized domain-classification tags in metadata
Correct Answer: D
Correct Answer
Answer D is correct because This option is correct because it normalizes business-domain labels across all ingested sources. That directly addresses source systems use inconsistent category names. Because source systems use inconsistent category names, this is the most direct design response.
Incorrect Answers
Answer A is incorrect because This option routes separate subject domains to independently tuned vector indexes. It fits when different corpora need independent tuning and lifecycle control. Here the decisive condition is source systems use inconsistent category names. In contrast, Store normalized domain-classification tags in metadata addresses that condition directly.
Answer B is incorrect because This option stores author identity as a consistent retrieval attribute. It fits when queries and audits must constrain results by approved author. Here the decisive condition is source systems use inconsistent category names. In contrast, Store normalized domain-classification tags in metadata addresses that condition directly.
Answer C is incorrect because This option applies caller-supplied structured conditions before ranking candidates. It fits when the client already knows the exact filter values. Here the decisive condition is source systems use inconsistent category names. In contrast, Store normalized domain-classification tags in metadata addresses that condition directly.
Question 7
The authoritative content lives in an internal wiki and changes frequently. The team wants the knowledge-base ingestion path tied to that system rather than maintaining manual export files. Which design choice best satisfies this requirement? Choose ONE.
- Set an S3 inclusion prefix for the approved path
- Use a CUSTOM data source with KnowledgeBaseDocuments APIs
- Use a knowledge-base connector for the internal wiki
- Trigger the vector update pipeline from source-change events
Correct Answer: C
Correct Answer
Answer C is correct because This option is correct because it connects the internal wiki directly to the ingestion path. That directly addresses the wiki is authoritative and manual exports should disappear. Because the wiki is authoritative and manual exports should disappear, this is the most direct design response.
Incorrect Answers
Answer A is incorrect because This option limits the connector to objects under an approved S3 prefix. It fits when only one governed folder belongs in the corpus. Here the decisive condition is the wiki is authoritative and manual exports should disappear. In contrast, Use a knowledge-base connector for the internal wiki addresses that condition directly.
Answer B is incorrect because This option accepts direct document ingest and delete calls with metadata. It fits when documents must be updated immediately without connector synchronization. Here the decisive condition is the wiki is authoritative and manual exports should disappear. In contrast, Use a knowledge-base connector for the internal wiki addresses that condition directly.
Answer D is incorrect because This option starts vector updates when source-change events occur. It fits when new procedures must become searchable within minutes. Here the decisive condition is the wiki is authoritative and manual exports should disappear. In contrast, Use a knowledge-base connector for the internal wiki addresses that condition directly.
Question 8
New incident procedures must become searchable within minutes of a source change. A nightly batch window is too slow. Which design choice best satisfies this requirement? Choose ONE.
- Use the metadata-only sync optimization
- Trigger the vector update pipeline from source-change events
- Batch embedding generation through a Lambda-based ingestion workflow
- Use a knowledge-base connector for the internal wiki
Correct Answer: B
Correct Answer
Answer B is correct because This option is correct because it starts vector updates when source-change events occur. That directly addresses new procedures must become searchable within minutes. Because new procedures must become searchable within minutes, this is the most direct design response.
Incorrect Answers
Answer A is incorrect because This option reuses stored embeddings while updating eligible metadata-only changes. It fits when document content is unchanged and embedding calls should be avoided. Here the decisive condition is new procedures must become searchable within minutes. In contrast, Trigger the vector update pipeline from source-change events addresses that condition directly.
Answer C is incorrect because This option groups large embedding workloads into controlled ingestion batches. It fits when millions of documents need efficient bulk embedding. Here the decisive condition is new procedures must become searchable within minutes. In contrast, Trigger the vector update pipeline from source-change events addresses that condition directly.
Answer D is incorrect because This option connects the internal wiki directly to the ingestion path. It fits when the wiki is authoritative and manual exports should disappear. Here the decisive condition is new procedures must become searchable within minutes. In contrast, Trigger the vector update pipeline from source-change events addresses that condition directly.
Question 9
The application already stores transactional records in Aurora PostgreSQL and needs vector similarity queries alongside relational joins and SQL-managed state. Which design choice best satisfies this requirement? Choose ONE.
- Use Amazon S3 Vectors for cost-focused, infrequent vector queries
- Use OpenSearch Serverless vector search
- Use Amazon Aurora PostgreSQL with pgvector for the vector store
- Scale the OpenSearch vector index with an appropriate sharding strategy
Correct Answer: C
Correct Answer
Answer C is correct because This option is correct because it keeps vectors beside relational records in Aurora PostgreSQL. That directly addresses vector search must coexist with SQL joins and transactions. Because vector search must coexist with SQL joins and transactions, this is the most direct design response.
Incorrect Answers
Answer A is incorrect because This option provides durable low-cost vector storage for sporadic query traffic. It fits when storage cost dominates and sustained throughput is modest. Here the decisive condition is vector search must coexist with SQL joins and transactions. In contrast, Use Amazon Aurora PostgreSQL with pgvector for the vector store addresses that condition directly.
Answer B is incorrect because This option provides managed OpenSearch vector search without cluster administration. It fits when the team wants OpenSearch-style search with serverless operations. Here the decisive condition is vector search must coexist with SQL joins and transactions. In contrast, Use Amazon Aurora PostgreSQL with pgvector for the vector store addresses that condition directly.
Answer D is incorrect because This option distributes index data and vector-query load across OpenSearch shards. It fits when one partition is the throughput bottleneck. Here the decisive condition is vector search must coexist with SQL joins and transactions. In contrast, Use Amazon Aurora PostgreSQL with pgvector for the vector store addresses that condition directly.
Question 10
A bulk corpus build must create embeddings for millions of objects efficiently, with controlled batching and orchestration around embedding calls. Which design choice best satisfies this requirement? Choose ONE.
- Batch embedding generation through a Lambda-based ingestion workflow
- Use the metadata-only sync optimization
- Choose a lower embedding dimensionality after recall benchmarking
- Trigger the vector update pipeline from source-change events
Correct Answer: A
Correct Answer
Answer A is correct because This option is correct because it groups large embedding workloads into controlled ingestion batches. That directly addresses millions of documents need efficient bulk embedding. Because millions of documents need efficient bulk embedding, this is the most direct design response.
Incorrect Answers
Answer B is incorrect because This option reuses stored embeddings while updating eligible metadata-only changes. It fits when document content is unchanged and embedding calls should be avoided. Here the decisive condition is millions of documents need efficient bulk embedding. In contrast, Batch embedding generation through a Lambda-based ingestion workflow addresses that condition directly.
Answer C is incorrect because This option uses a smaller vector representation after recall testing validates it. It fits when storage and latency matter more once recall remains acceptable. Here the decisive condition is millions of documents need efficient bulk embedding. In contrast, Batch embedding generation through a Lambda-based ingestion workflow addresses that condition directly.
Answer D is incorrect because This option starts vector updates when source-change events occur. It fits when new procedures must become searchable within minutes. Here the decisive condition is millions of documents need efficient bulk embedding. In contrast, Batch embedding generation through a Lambda-based ingestion workflow addresses that condition directly.
Question 11
Benchmarks show a smaller embedding representation preserves the required recall, while the current vectors are driving excessive storage and transfer cost. Which design choice best satisfies this requirement? Choose ONE.
- Re-embed the corpus with the new embedding model
- Choose a lower embedding dimensionality after recall benchmarking
- Batch embedding generation through a Lambda-based ingestion workflow
- Monitor KNNGraphMemoryUsage and size OpenSearch data nodes for graph memory
Correct Answer: B
Correct Answer
Answer B is correct because This option is correct because it uses a smaller vector representation after recall testing validates it. That directly addresses storage and latency matter more once recall remains acceptable. Because storage and latency matter more once recall remains acceptable, this is the most direct design response.
Incorrect Answers
Answer A is incorrect because This option regenerates stored vectors in the new embedding model’s representation space. It fits when an embedding-model migration changes vector geometry. Here the decisive condition is storage and latency matter more once recall remains acceptable. In contrast, Choose a lower embedding dimensionality after recall benchmarking addresses that condition directly.
Answer C is incorrect because This option groups large embedding workloads into controlled ingestion batches. It fits when millions of documents need efficient bulk embedding. Here the decisive condition is storage and latency matter more once recall remains acceptable. In contrast, Choose a lower embedding dimensionality after recall benchmarking addresses that condition directly.
Answer D is incorrect because This option tracks OpenSearch k-NN graph memory against its circuit-breaker allowance. It fits when ANN graph memory pressure is causing vector-search instability. Here the decisive condition is storage and latency matter more once recall remains acceptable. In contrast, Choose a lower embedding dimensionality after recall benchmarking addresses that condition directly.
Question 12
Users submit terse product acronyms and shorthand that often fail to match the richer language used in documents. The team wants to enrich the request before retrieval. Which design choice best satisfies this requirement? Choose ONE.
- Enable query decomposition
- Apply query expansion before retrieval
- Configure implicit filtering from the user query
- Use Step Functions for a deterministic query-transformation workflow
Correct Answer: B
Correct Answer
Answer B is correct because This option is correct because it adds related terms or context before retrieval. That directly addresses terse acronym-heavy questions lack enough matching signals. Because terse acronym-heavy questions lack enough matching signals, this is the most direct design response.
Incorrect Answers
Answer A is incorrect because This option splits one compound request into simpler retrieval subqueries. It fits when a single request contains several independent information needs. Here the decisive condition is terse acronym-heavy questions lack enough matching signals. In contrast, Apply query expansion before retrieval addresses that condition directly.
Answer C is incorrect because This option derives metadata restrictions from natural-language query context. It fits when users express filters in prose rather than structured fields. Here the decisive condition is terse acronym-heavy questions lack enough matching signals. In contrast, Apply query expansion before retrieval addresses that condition directly.
Answer D is incorrect because This option orchestrates deterministic query normalization and routing with explicit workflow state. It fits when query rewrites require governed ordering and auditability. Here the decisive condition is terse acronym-heavy questions lack enough matching signals. In contrast, Apply query expansion before retrieval addresses that condition directly.
Question 13
Queries must pass through governed normalization, classification, and routing steps in a repeatable order with explicit state and auditability. Which design choice best satisfies this requirement? Choose ONE.
- Use agentic retrieval
- Apply query expansion before retrieval
- Publish a stable retrieval API contract between the FM layer and vector backend
- Use Step Functions for a deterministic query-transformation workflow
Correct Answer: D
Correct Answer
Answer D is correct because This option is correct because it orchestrates deterministic query normalization and routing with explicit workflow state. That directly addresses query rewrites require governed ordering and auditability. Because query rewrites require governed ordering and auditability, this is the most direct design response.
Incorrect Answers
Answer A is incorrect because This option plans iterative searches and continues until evidence is sufficient. It fits when a multi-hop question needs more than one planned retrieval step. Here the decisive condition is query rewrites require governed ordering and auditability. In contrast, Use Step Functions for a deterministic query-transformation workflow addresses that condition directly.
Answer B is incorrect because This option adds related terms or context before retrieval. It fits when terse acronym-heavy questions lack enough matching signals. Here the decisive condition is query rewrites require governed ordering and auditability. In contrast, Use Step Functions for a deterministic query-transformation workflow addresses that condition directly.
Answer C is incorrect because This option hides vector-store details behind a stable retrieval contract. It fits when applications must survive a later backend replacement. Here the decisive condition is query rewrites require governed ordering and auditability. In contrast, Use Step Functions for a deterministic query-transformation workflow addresses that condition directly.
Question 14
A foundation model should decide during its reasoning whether it needs evidence, then invoke the vector-search capability through a defined tool schema. Which design choice best satisfies this requirement? Choose ONE.
- Use an MCP client for standardized vector-query tool access
- Expose vector retrieval through foundation-model function calling
- Call the Retrieve API
- Publish a stable retrieval API contract between the FM layer and vector backend
Correct Answer: B
Correct Answer
Answer B is correct because This option is correct because it exposes retrieval as a callable tool with a defined schema. That directly addresses the model should decide when evidence retrieval is needed. Because the model should decide when evidence retrieval is needed, this is the most direct design response.
Incorrect Answers
Answer A is incorrect because This option invokes retrieval through the Model Context Protocol. It fits when multiple model runtimes require an interoperable tool interface. Here the decisive condition is the model should decide when evidence retrieval is needed. In contrast, Expose vector retrieval through foundation-model function calling addresses that condition directly.
Answer C is incorrect because This option returns retrieved passages and metadata without generating the final answer. It fits when the application owns downstream ranking or generation. Here the decisive condition is the model should decide when evidence retrieval is needed. In contrast, Expose vector retrieval through foundation-model function calling addresses that condition directly.
Answer D is incorrect because This option hides vector-store details behind a stable retrieval contract. It fits when applications must survive a later backend replacement. Here the decisive condition is the model should decide when evidence retrieval is needed. In contrast, Expose vector retrieval through foundation-model function calling addresses that condition directly.
Question 15
A platform standard requires retrieval tools to be consumable through Model Context Protocol by multiple compatible model runtimes. Which design choice best satisfies this requirement? Choose ONE.
- Use an MCP client for standardized vector-query tool access
- Expose vector retrieval through foundation-model function calling
- Publish a stable retrieval API contract between the FM layer and vector backend
- Use a knowledge-base connector for the internal wiki
Correct Answer: A
Correct Answer
Answer A is correct because This option is correct because it invokes retrieval through the Model Context Protocol. That directly addresses multiple model runtimes require an interoperable tool interface. Because multiple model runtimes require an interoperable tool interface, this is the most direct design response.
Incorrect Answers
Answer B is incorrect because This option exposes retrieval as a callable tool with a defined schema. It fits when the model should decide when evidence retrieval is needed. Here the decisive condition is multiple model runtimes require an interoperable tool interface. In contrast, Use an MCP client for standardized vector-query tool access addresses that condition directly.
Answer C is incorrect because This option hides vector-store details behind a stable retrieval contract. It fits when applications must survive a later backend replacement. Here the decisive condition is multiple model runtimes require an interoperable tool interface. In contrast, Use an MCP client for standardized vector-query tool access addresses that condition directly.
Answer D is incorrect because This option connects the internal wiki directly to the ingestion path. It fits when the wiki is authoritative and manual exports should disappear. Here the decisive condition is multiple model runtimes require an interoperable tool interface. In contrast, Use an MCP client for standardized vector-query tool access addresses that condition directly.
Question 16
Several applications consume retrieval, but the backend vector store may be replaced later. Consumers must not depend on vendor-specific request formats. Which design choice best satisfies this requirement? Choose ONE.
- Use an MCP client for standardized vector-query tool access
- Expose vector retrieval through foundation-model function calling
- Publish a stable retrieval API contract between the FM layer and vector backend
- Route specialized domains to separate vector indexes
Correct Answer: C
Correct Answer
Answer C is correct because This option is correct because it hides vector-store details behind a stable retrieval contract. That directly addresses applications must survive a later backend replacement. Because applications must survive a later backend replacement, this is the most direct design response.
Incorrect Answers
Answer A is incorrect because This option invokes retrieval through the Model Context Protocol. It fits when multiple model runtimes require an interoperable tool interface. Here the decisive condition is applications must survive a later backend replacement. In contrast, Publish a stable retrieval API contract between the FM layer and vector backend addresses that condition directly.
Answer B is incorrect because This option exposes retrieval as a callable tool with a defined schema. It fits when the model should decide when evidence retrieval is needed. Here the decisive condition is applications must survive a later backend replacement. In contrast, Publish a stable retrieval API contract between the FM layer and vector backend addresses that condition directly.
Answer D is incorrect because This option routes separate subject domains to independently tuned vector indexes. It fits when different corpora need independent tuning and lifecycle control. Here the decisive condition is applications must survive a later backend replacement. In contrast, Publish a stable retrieval API contract between the FM layer and vector backend addresses that condition directly.
Question 17
The team wants to force HYBRID retrieval in Bedrock Knowledge Bases on OpenSearch Serverless, but the current index lacks the text capability required for that override. Which design choice best satisfies this requirement? Choose ONE.
- Provision a filterable text field before forcing HYBRID search in OpenSearch Serverless
- Use OpenSearch Serverless vector search
- Apply query expansion before retrieval
- Apply an explicit metadata filter
Correct Answer: A
Correct Answer
Answer A is correct because This option is correct because it adds the filterable text capability required before HYBRID override works. That directly addresses Bedrock must force HYBRID search on OpenSearch Serverless. Because Bedrock must force HYBRID search on OpenSearch Serverless, this is the most direct design response.
Incorrect Answers
Answer B is incorrect because This option provides managed OpenSearch vector search without cluster administration. It fits when the team wants OpenSearch-style search with serverless operations. Here the decisive condition is Bedrock must force HYBRID search on OpenSearch Serverless. In contrast, Provision a filterable text field before forcing HYBRID search in OpenSearch Serverless addresses that condition directly.
Answer C is incorrect because This option adds related terms or context before retrieval. It fits when terse acronym-heavy questions lack enough matching signals. Here the decisive condition is Bedrock must force HYBRID search on OpenSearch Serverless. In contrast, Provision a filterable text field before forcing HYBRID search in OpenSearch Serverless addresses that condition directly.
Answer D is incorrect because This option applies caller-supplied structured conditions before ranking candidates. It fits when the client already knows the exact filter values. Here the decisive condition is Bedrock must force HYBRID search on OpenSearch Serverless. In contrast, Provision a filterable text field before forcing HYBRID search in OpenSearch Serverless addresses that condition directly.
Question 18
A knowledge base uses custom embeddings, and the team still wants a second-stage relevance model. The managed reranking option for managed embeddings is not applicable. Which design choice best satisfies this requirement? Choose ONE.
- Rerank retrieved candidates
- Choose a lower embedding dimensionality after recall benchmarking
- Call the Retrieve API
- Use a custom reranker for a knowledge base with custom embeddings
Correct Answer: D
Correct Answer
Answer D is correct because This option is correct because it applies a separately configured relevance model after retrieval. That directly addresses custom embeddings still require second-stage relevance scoring. Because custom embeddings still require second-stage relevance scoring, this is the most direct design response.
Incorrect Answers
Answer A is incorrect because This option rescales a retrieved candidate set with a stronger relevance model. It fits when initial recall is good but final ordering is weak. Here the decisive condition is custom embeddings still require second-stage relevance scoring. In contrast, Use a custom reranker for a knowledge base with custom embeddings addresses that condition directly.
Answer B is incorrect because This option uses a smaller vector representation after recall testing validates it. It fits when storage and latency matter more once recall remains acceptable. Here the decisive condition is custom embeddings still require second-stage relevance scoring. In contrast, Use a custom reranker for a knowledge base with custom embeddings addresses that condition directly.
Answer C is incorrect because This option returns retrieved passages and metadata without generating the final answer. It fits when the application owns downstream ranking or generation. Here the decisive condition is custom embeddings still require second-stage relevance scoring. In contrast, Use a custom reranker for a knowledge base with custom embeddings addresses that condition directly.
Question 19
A very large vector corpus is queried only occasionally. The main goal is low-cost, durable, elastic vector storage rather than maximum sustained query throughput. Which design choice best satisfies this requirement? Choose ONE.
- Use Amazon S3 Vectors for cost-focused, infrequent vector queries
- Use Amazon Aurora PostgreSQL with pgvector for the vector store
- Scale the OpenSearch vector index with an appropriate sharding strategy
- Use OpenSearch Serverless vector search
Correct Answer: A
Correct Answer
Answer A is correct because This option is correct because it provides durable low-cost vector storage for sporadic query traffic. That directly addresses storage cost dominates and sustained throughput is modest. Because storage cost dominates and sustained throughput is modest, this is the most direct design response.
Incorrect Answers
Answer B is incorrect because This option keeps vectors beside relational records in Aurora PostgreSQL. It fits when vector search must coexist with SQL joins and transactions. Here the decisive condition is storage cost dominates and sustained throughput is modest. In contrast, Use Amazon S3 Vectors for cost-focused, infrequent vector queries addresses that condition directly.
Answer C is incorrect because This option distributes index data and vector-query load across OpenSearch shards. It fits when one partition is the throughput bottleneck. Here the decisive condition is storage cost dominates and sustained throughput is modest. In contrast, Use Amazon S3 Vectors for cost-focused, infrequent vector queries addresses that condition directly.
Answer D is incorrect because This option provides managed OpenSearch vector search without cluster administration. It fits when the team wants OpenSearch-style search with serverless operations. Here the decisive condition is storage cost dominates and sustained throughput is modest. In contrast, Use Amazon S3 Vectors for cost-focused, infrequent vector queries addresses that condition directly.
Question 20
An OpenSearch vector workload becomes unstable as approximate-nearest-neighbor graph memory approaches its circuit-breaker allowance. The team needs the operational signal that directly tracks this pressure. Which design choice best satisfies this requirement? Choose ONE.
- Scale the OpenSearch vector index with an appropriate sharding strategy
- Monitor KNNGraphMemoryUsage and size OpenSearch data nodes for graph memory
- Use approximate k-NN with HNSW for the large vector index
- Choose a lower embedding dimensionality after recall benchmarking
Correct Answer: B
Correct Answer
Answer B is correct because This option is correct because it tracks OpenSearch k-NN graph memory against its circuit-breaker allowance. That directly addresses ANN graph memory pressure is causing vector-search instability. Because ANN graph memory pressure is causing vector-search instability, this is the most direct design response.
Incorrect Answers
Answer A is incorrect because This option distributes index data and vector-query load across OpenSearch shards. It fits when one partition is the throughput bottleneck. Here the decisive condition is ANN graph memory pressure is causing vector-search instability. In contrast, Monitor KNNGraphMemoryUsage and size OpenSearch data nodes for graph memory addresses that condition directly.
Answer C is incorrect because This option uses approximate HNSW search to reduce nearest-neighbor latency at scale. It fits when exact vector search is too slow for the corpus size. Here the decisive condition is ANN graph memory pressure is causing vector-search instability. In contrast, Monitor KNNGraphMemoryUsage and size OpenSearch data nodes for graph memory addresses that condition directly.
Answer D is incorrect because This option uses a smaller vector representation after recall testing validates it. It fits when storage and latency matter more once recall remains acceptable. Here the decisive condition is ANN graph memory pressure is causing vector-search instability. In contrast, Monitor KNNGraphMemoryUsage and size OpenSearch data nodes for graph memory addresses that condition directly.
