What You Need to Know
A retrieval strategy is how an assistant finds the information that an answer depends on. The choice of strategy depends on two factors. Data shape is the form the information takes. Query pattern is how people ask for it. Then ask which failure matters most: staleness, missed recall (a relevant document that retrieval does not find), wrong precision or latency. Pick the cheapest strategy whose failure mode is acceptable. Objective 3.5 designed one pipeline. This objective decides when that pipeline is the right choice.
This lesson uses one made-up example throughout: an energy supplier whose customer-service assistant reads a live billing system, policies, price-plan guides and code. One system routinely needs several strategies at once.
The Shape Axis
Data shape decides which source holds the true data.
Live structured data, such as databases, catalogues and account records, already gives precise, current answers through its own query interface. The right integration is a query tool against the system of record, which is the authoritative source of the data. Examples are an API tool and text-to-SQL, which turns a question into a database query. Embedding its rows produces a stale copy. The copy is wrong as soon as it is indexed, and it has lost the structure that made queries precise. The supplier's assistant fetches a customer's account balance this way.
Narrative documents, such as policies, manuals and articles, have no query language. Semantic retrieval over well-chunked text fits them, which is the 3.5 pipeline's strongest case. The supplier's late-payment policy belongs here.
Code and rigidly structured text, such as the supplier's bill-calculation code, are between these two shapes. Similarity search alone fails on exact symbols and cross-references. So retrieval follows the structure: search, open, follow references.
The Query-Pattern Axis
The same corpus is used in different ways for different questions:
| Query pattern | Cue in the question | Native strategy | Energy supplier example |
|---|---|---|---|
| Exact lookup | An identifier: part number, error code, clause reference | Keyword or filtered search against the right source | A customer account by account number |
| Conceptual | "Summarise", "explain", "what is our policy on" | Semantic retrieval over chunked text | "What is our policy on late payment?" |
| Multi-hop / exploratory | Comparison across documents, unknown search path | Agentic retrieval: search, read, refine, search again | Cancellation terms compared across three price-plan guides |
| Recall-critical | "Must not miss a relevant document" (compliance, legal) | Hybrid search plus reranking | A compliance audit that must find every complaint about a wrong bill |
Hybrid search combines embedding search with keyword search. Reranking is a later step that scores the retrieved chunks for relevance and keeps the best ones.
Anthropic's context-engineering guidance describes the multi-hop case as just-in-time retrieval. The agent keeps lightweight identifiers and loads data as questions arise. It discovers what matters progressively, instead of loading a guessed set of data at the start.
When Agentic Retrieval Earns Its Price
Agentic retrieval spends latency and tokens to get precision on questions that no single query answers. This cost is justified for exploratory and comparative work. It is unjustified for a lookup that a metadata filter would answer in one step. A metadata filter narrows a search by a stored field, such as date or source.
The failure that matters decides which strategy to use. Recall-critical corpora justify hybrid search plus reranking, because a missed document is the expensive failure. Lookups with a latency budget (a limit on how long the answer may take) justify a single query, because for them a slow answer is the expensive failure. The supplier's three-guide comparison justifies the agentic loop. Its account lookup does not.
Hybrid Is a Design Answer, Not a Compromise
One system can combine a system-of-record tool for live lookups, semantic retrieval over manuals and agentic retrieval for comparisons. The energy supplier uses all three. A small, stable, high-value corpus needs no retrieval at all. The cached prompt contains the whole corpus.
The Selection Method
Work through four steps:
- Name the data shape.
- Name the query pattern.
- Name the failure that matters most.
- Pick the cheapest strategy whose failure mode is acceptable.
For the supplier's account balance, the data is live and structured, and the query is an exact lookup. Staleness matters, so use a query tool.
In the exam
A question that presents three query families tests whether you match each family to the strategy that fits it best or force one strategy onto all three. Answers that combine strategies are correct more often than answers that use one strategy. The selection method eliminates the wrong options by a fixed procedure. An option that embeds live operational data fails on staleness before you examine any other property. An impressive-looking option that uses agentic retrieval on a simple lookup fails on cost. The question never justified that cost. Domain 2 covers the cached-prompt boundary.
Exam Traps
Embedding live operational data, such as stock levels, balances and account records, into a vector store
Stale the moment it is indexed, and the structure that made queries precise is gone. The database already answers exactly and currently through a query tool, so use that.
Applying the retrieval approach that worked for the first corpus to every data shape that follows
Each shape has one strategy that fits it best. A strategy reused from another shape brings the wrong failure mode (stale rows, missed identifiers, weak recall) into a corpus that did not have it.
Deploying agentic multi-hop retrieval on a simple, latency-budgeted lookup
Latency and tokens are spent to get precision a simple lookup never needed. A single query with a metadata filter answers it in one hop.
Choosing a precision- or latency-first option where the question says no relevant document may be missed
In recall-critical work, a gap in coverage is the expensive failure. Use hybrid search plus reranking.
Practice Scenario
An enterprise assistant must answer three query families: the current stock level of part K-4402 (live ERP data), a summary of the data-retention policy (a 300-page manual), and how clause 14 has changed across the last three supplier contracts (an archive of several thousand contracts). Staff ask policy questions in everyday words that rarely match the manual's terms. Which retrieval design fits?
Build Exercise
Retrieval Strategy Matrix for One Organisation's Data Estate
What you'll learn
- How to classify data sources by shape and queries by pattern
- How to name the failure that matters per source-pattern pair
- How to assign the cheapest strategy whose failure mode is acceptable
- How to justify an expensive assignment by naming exactly what breaks with a cheaper strategy
- List the data sources one assistant must serve and classify each by shape: live structured, narrative documents, code or rigid text, small-and-stable
Why: Shape decides which source holds the true data, and the classification is the first step that eliminates candidate strategies.
You should see: A source list with one shape label each — and at least two different shapes, or the exercise needs more varied data sources.
- For each source, list the real query patterns against it: exact lookup, conceptual, multi-hop, recall-critical
Why: The same corpus is used in different ways for different questions — each strategy is chosen for a query pattern, not for a source.
You should see: A source-times-pattern grid with the cells that actually occur marked.
- Name the failure that matters most per source-pattern pair: staleness, missed recall (a relevant document that retrieval does not find), wrong precision, or latency
Why: The failure that matters most decides between candidate strategies — it turns a preference into a reasoned choice.
You should see: One named failure per marked cell in the grid.
- Assign each pair the cheapest strategy whose failure mode is acceptable: system-of-record tool, semantic retrieval, hybrid plus reranking, agentic retrieval, or cached prompt with no retrieval
Why: This is the selection method the exam tests, applied to all sources: cheapest good-enough option, never chosen to impress.
You should see: A completed matrix, each cell holding a strategy and implicitly an argument.
- Defend the two most expensive assignments: what specifically breaks if the cheaper strategy is used, and how would the failure be detected
Why: An expensive strategy that cannot name its breakage scenario is spending to look impressive — and detection makes the choice part of monitoring.
You should see: Two short defences, each naming the concrete failure and the observability signal (objective 3.4) that would catch it.
Sources
- Effective context engineering for AI agents — Anthropic
- Introducing Contextual Retrieval — Anthropic
- Claude Certified Architect Professional Exam Guide (v1.0) — Domain 3 — Anthropic