AWS Adds Single-Shot RAG and Zero-LLM Retrieval to Bedrock Knowledge Bases

Summary:Amazon Bedrock Managed Knowledge Base now exposes three retrieval modes through the same AgenticRetrieveStream API. With maxAgentIteration set to 0, it returns deduplicated passages without invoking a foundation model. A value of 1 performs one retrieval and one grounded answer-generation call, while values from 2 to 5 enable query decomposition, evidence evaluation, and iterative searches before generation. The update gives applications a unified way to serve simple lookups, complex comparison

---

One API for Three Retrieval Paths

Amazon Web Services has expanded AgenticRetrieveStream in Amazon Bedrock Managed Knowledge Base with single-shot RAG and retrieval-only requests. The maxAgentIteration parameter now selects among three execution paths:

• maxAgentIteration=0: retrieve once and return deduplicated passages without invoking a foundation model;
• maxAgentIteration=1: retrieve once and make one grounded answer-generation call;
• maxAgentIteration=2..5: perform multiple planning, query-rewrite, evidence-evaluation, and retrieval rounds before generating an answer.

Clients no longer need separate application paths for simple retrieval, basic RAG, and agentic retrieval. However, the service does not automatically classify request complexity; the application must choose the mode.

How Iterations Are Counted

AWS counts one foundation-model call as one iteration. A value of n permits up to n-1 agentic planning rounds, with the final iteration reserved for answer generation. Each planning round consists of a query rewrite and a retrieval operation.

Consequently, maxAgentIteration=1 performs no agentic planning: it retrieves once and generates once. A value of 4 allows up to three rounds of planning and retrieval before the final generation call. The parameter therefore controls model work rather than simply the number of vector searches.

Speculative retrieval does not count against the limit. It runs ahead of the first planning step to reduce initial response latency, but its effectiveness still depends on later ranking and query refinement.

Engineering Behavior of Each Mode

Retrieval Only

With maxAgentIteration=0 and generateResponse=false, the service returns retrieved passages without any foundation-model call. This mode suits source review, audit workflows, and applications that feed results into deterministic code.

Setting generateResponse=true at zero iterations returns a 4xx error because answer generation necessarily requires a model call. The behavior resembles the Retrieve API, but the response is not completely identical: passages arrive through the stream's result event without the score or location fields returned by the direct API.

Single-Shot RAG

maxAgentIteration=1 implements the classic retrieve-once, generate-once pattern without query rewriting or agentic planning. It fits direct questions whose answer is contained in a well-scoped checklist or document.

Compared with multi-round agentic retrieval, it reduces planning calls and repeated searches, making latency and cost easier to predict. The trade-off is that the system cannot proactively recover evidence spread across multiple documents. The announcement's banking example is illustrative, not a measured customer deployment with independently reported accuracy or cost figures.

Agentic Retrieval

At values from 2 through 5, a foundation model decomposes the question, creates targeted queries, and evaluates whether the retrieved evidence is sufficient. If it detects gaps, it can rewrite the query and search additional knowledge bases before producing a cited answer.

This mode is better suited to comparative, conditional, exploratory, and cross-knowledge-base questions. It also adds model calls, retrieval work, and reranking overhead. A higher limit does not guarantee a better answer: faulty decomposition may direct retrieval toward irrelevant material, while additional rounds can increase latency, cost, and error accumulation.

Balancing Latency, Cost, and Evidence Quality

| Mode | Model calls | Best suited for | Main limitation | |---|---:|---|---| | Retrieval only | 0 | Source review and deterministic consumption | No synthesized answer; ranking errors are exposed directly | | Single-shot RAG | 1 generation | Direct, tightly scoped queries | One retrieval can miss cross-document evidence | | Agentic retrieval | Multiple planning calls plus generation | Comparisons and cross-corpus questions | Higher tail latency and cost; weak planning can amplify errors |

The main value of the change is not a claimed accuracy gain for a single configuration. It gives application teams explicit control over the retrieval budget and removes the need for separate client paths. A production system should classify requests and assign each route the lowest iteration limit that meets its quality target.

The Banking Example

AWS illustrates the feature with a fictional internal banking assistant:

• For the documents required to open a business checking account, it uses maxAgentIteration=1 and generateResponse=true.
• For a comparison covering sole proprietors, LLCs, and nonresident owners, it uses maxAgentIteration=4, leaving room for three supplementary retrieval rounds.
• For displaying policy source text, it uses maxAgentIteration=0 and generateResponse=false.

The example demonstrates application-level routing, but AWS states that it is not a customer deployment or a measured result. The assistant retrieves and explains policy; it does not approve accounts or replace human review and authorization.

Deployment Guidance

Teams should not simply move every request to the highest iteration value. A more reliable migration process is:

1. Sample real requests covering single-document lookup, cross-document comparison, missing evidence, and permission boundaries.
2. Compare single-shot RAG with the existing agentic configuration using the same knowledge base and generation settings.
3. Measure retrieval quality, answer correctness, citation support, P50 and P95 latency, model-call count, and per-request cost.
4. Assign the lowest iteration limit that satisfies each workflow's quality requirement.
5. Keep test cases whose answers do not exist in the corpus, preventing a fast response from hiding an evidence gap.
6. Audit citations, access controls, and generated answers separately; displaying retrieved passages alone does not establish a production-grade evidence chain.

The expanded range preserves existing behavior for values from 2 through 5. Organizations can therefore introduce zero-iteration retrieval for source-oriented workflows and single-shot RAG for simple traffic before adjusting more complex requests.

Compatibility Boundaries

All three modes support multiple knowledge bases. Single-shot retrieval does not mean that only one knowledge base can be queried. The update changes planning and generation budgets, not the number of available knowledge sources.

Applications must still choose the mode and control answer generation with generateResponse. Retrieval-only mode removes model latency and generation cost, but it also removes the natural-language synthesis layer. Callers then need to handle ranking, source location, and structured output themselves.

---

Source: AWS Machine Learning Blog (2026-11-09)

— END —

Tags: #Amazon Bedrock Follow #Agentic RAG Follow #Knowledge Base Retrieval Follow #AgenticRetrieveStream Follow