Semantic, hybrid search and reranking: AIP-C01 study notes and practice questions
What to know about Semantic, hybrid search and reranking for the AIP-C01 exam, from the Foundation Model Integration, Data Management, and Compliance domain (31% of the exam guide), then 1 free question to test it. Part of the AIP-C01 study guide.
AIP-C01 asks how to improve retrieval when pure vector search misses exact terms or returns weak matches, and where a reranker fits in the pipeline.
Why each method helps
Semantic search matches meaning, so “laptop battery replacement” finds “notebook power cell” guides. It fails on exact tokens such as product codes, error IDs and rare names, where the embedding may not separate similar strings.
Keyword search matches exact terms. It fails when users use synonyms or paraphrase.
Hybrid search runs both and merges the scores. It is the right default for mixed corpora. A managed Amazon Bedrock knowledge base always uses hybrid search. A knowledge base you configure uses semantic search by default, and you can set the search type to hybrid in the Retrieve request when the store supports it, such as OpenSearch or Aurora PostgreSQL. In your own OpenSearch application, neural search can call an embedding model through a Bedrock connector at query time, so the query and the index use the same model.
The reranking pipeline
Retrieve a larger set of candidates, for example 20 to 50 chunks, with hybrid search.
Send the query and each candidate to a reranker model, such as Amazon Rerank or a Cohere Rerank model, which scores relevance more precisely than the first-stage search.
Keep only the top few chunks for the prompt.
Retrieving more than you pass to the model gives the reranker room to promote a relevant chunk that the first search ranked low. The reranker adds latency and cost per query, so measure whether it earns that cost on your test questions.
Score thresholds
A score threshold drops chunks below a relevance level, so the model is not given weak context to answer from. Set it from evaluation data. Too high and the system answers “I don’t know” to real questions. Too low and noise reaches the prompt.
How to choose
Scenario wording
Choose
“Customers search by exact product codes and by description”
Hybrid search
“Right chunk is retrieved but ranked too low”
Reranker over a larger candidate set
“Answers include loosely related text when nothing relevant exists”
Score threshold
“Keyword search misses synonyms”
Add semantic search (hybrid)
“Lowest added latency, good enough relevance”
Hybrid search without a reranker
Common traps
Keyword-only search for a question built on paraphrases, which a semantic index would catch.
Vector-only search for exact codes, where the right item may not rank first.
Reranking only the final few results, which gives the reranker nothing to fix.
Retrieving exactly the number you pass to the model, which removes the benefit of reranking.
Adding a reranker to fix bad chunking or a wrong embedding model. Fix the root cause first.
A score threshold set once and never tested, which silently hides relevant content after a model change.
A hardware distributor runs a search box over 400,000 product records. Buyers type exact part numbers, such as XK-4471-B, and also describe what they need in plain language, such as a quiet pump for well water.
Vector-only search finds the descriptions well but often returns the wrong item for a part number, because codes carry little meaning in an embedding. Keyword-only search finds the part numbers but misses the descriptions.
The results must satisfy both query types in one search box, with the best match at the top, and the distributor wants to keep its current OpenSearch index.
Which change meets these requirements?
Choose one answer.
AA. Run hybrid search in the OpenSearch index, combining keyword matching on part numbers with vector similarity on descriptions, and then rerank the merged results.
BB. Keep vector-only search, and increase the chunk size of each product record so that the part number sits closer to the surrounding words in the embedding.
CC. Keep keyword-only search, and maintain a long synonym list that maps common descriptions such as quiet pump or well water to the matching part numbers.
DD. Fine-tune an embedding model on pairs of part numbers and descriptions so codes land near their descriptions, and keep vector-only search with it.
B. Wrong: Longer records make the embedding more diluted, and the part number still carries little meaning in a vector. This does not fix exact code matches, so the first requirement is missed.
C. Wrong: A synonym list has to be updated by hand for every new product and phrasing, and it can never cover the variety of plain-language descriptions. It misses the meaning requirement and adds ongoing maintenance.
D. Wrong: Fine-tuning needs a large set of labeled pairs and a training project, which delays the change and adds cost. Codes may still match poorly against unseen numbers, so this does not guarantee exact matches.
Explanation
Hybrid search covers exact part numbers through keywords and plain-language descriptions through vectors, within the existing OpenSearch index, and a reranker sets the final order. Longer records, synonym lists and fine-tuning each leave one query type poorly served or add ongoing effort.
Want more on Semantic, hybrid search and reranking? The 5 full-length practice sets have 5 more Semantic, hybrid search and reranking questions, 375 in all, none of them in the free exam. One payment of $14.99, no subscription. See the practice sets →