AI 产品经理04 Architecture & Risk
Chapter 05
5 / 10

RAG & Agent Strategy: Complex System Design

⏱️ 75 min

Master RAG architecture design, multi-agent collaboration patterns, and Function Calling standards to build enterprise-grade intelligent systems

CHAPTER DECISION GOAL
01Product question

Master RAG architecture design, multi-agent collaboration patterns, and Function Calling standards to build enterprise-grade intelligent systems

02Reviewable evidence

A reviewable product artefact: an assumption, prototype, evaluation result or launch decision.

03Definition of done

State the decision, the supporting evidence and the gate for moving to the next stage.

RAG and Agent are the two directions where AI PMs are most easily led astray by buzzwords. Many roadmaps open with "build an Agent" or "add RAG," but the user task hasn't even been defined yet. Result: lots of architecture terminology, very little product value.

This page isn't about implementation details. It's about judging from a PM perspective: when to use RAG, when to use Agent, and when neither should be rushed.

RAG Agent Strategy Map


Bottom Line: Ask About the User Task First, Then Choose System Form

A more stable decision sequence:

  1. Does the user need "more accurate answers" or "more complex execution"
  2. Are you lacking knowledge access capability or task orchestration capability
  3. Are risk and cost worth introducing a more complex system

If these three steps aren't thought through first, teams easily mistake technical complexity for product progress.


What Problems Are Better Suited for RAG

RAG's core value isn't making answers "smarter." It's making answers more grounded.

It fits better when:

ScenarioWhy it fits
Internal knowledge Q&ANeeds to reference company docs and rules
Help center / support copilotNeeds to answer based on existing knowledge
Policy, process, product docs retrievalNeeds source-backed answers
Long document Q&AThe model doesn't inherently know your private content

If the problem is fundamentally "the model doesn't know this material," RAG is usually the right direction.


What Problems Are Better Suited for Agent

Agent's core value isn't being "more human-like." It's being able to execute multi-step tasks.

It fits better when:

ScenarioWhy it fits
Multi-step research workflowNeeds to search, organize, generate results
Tasks requiring multiple tool callse.g., query data, write report, send notification
Complex operation workflowsNeeds to judge next actions
Semi-automated task executionNot just answering -- actually doing things

If the task is fundamentally "do A first, then judge, then do B," Agent starts to make sense.


When You Shouldn't Rush to Add RAG

SituationReason
Document quality is poorGarbage in, garbage answers out
Knowledge update process isn't establishedWill go stale quickly
What users actually want isn't Q&AYou might be optimizing the wrong problem
Team hasn't defined source trustRetrieved doesn't mean usable

Many cases of "RAG performing poorly" aren't actually model problems. They're knowledge base governance failures.


When You Shouldn't Casually Add Agent

SituationReason
Task steps are actually fixedRegular workflow automation might be more stable
One wrong step is very costlyAgent's autonomy amplifies risk
Single-step quality isn't solved yetMulti-step chains only amplify problems
Users don't actually need autonomyYou're adding complexity, not value

Agent isn't an upgraded chat box. Many products actually only need a clear flow + a few tool calls, not a full agent loop.


A More Practical Decision Framework

Start with this table:

Problem typeMore likely solution
Lacking knowledgeRAG
Lacking step executionAgent
Lacking bothRAG + Agent, but layer them first
Just a regular form or rule flowMight not need AI system complexity at all

What PMs should avoid most is choosing the most complex solution because "it sounds more advanced."


RAG PM Focus Points: Beyond Retrieval Accuracy

More important questions to watch:

Decision pointWhy PMs should care
Source coverageDoes the knowledge base cover what users will ask
Update freshnessHow often is knowledge refreshed
Citation UXCan users see the source
Failure handlingWhat happens when nothing is found
Trust boundaryWhich sources can be trusted

Whether RAG works well depends heavily on knowledge operations, not just per-retrieval metrics.


Agent PM Focus Point: Controllability

The question Agents should face isn't "is it cool" but:

  1. Will it call tools it shouldn't call
  2. Will it keep executing based on wrong premises
  3. Is each step observable
  4. Can it abort or escalate to human on failure

If these can't be answered clearly, the Agent approach usually isn't mature enough.


When RAG + Agent Ship Together, Layer Them First

A more stable approach isn't building one "big comprehensive Agent" at once. Break it into:

knowledge layer -> retrieval layer -> decision layer -> action layer

This way you can distinguish:

  • Was it a retrieval error
  • Was it a judgment error
  • Was it a tool execution error

Once systems get complex, the worst thing is having nobody know which layer the error happened in.


Most Overlooked Costs

RAG and Agent costs aren't limited to API bills.

They also include:

  • Document governance costs
  • Eval and monitoring costs
  • Prompt / workflow maintenance costs
  • Bad case handling costs

If PMs only budget "model call costs," they'll typically underestimate significantly.


A Sufficient Strategy Review Question Set

Before discussing RAG or Agent, have the team answer:

  1. What's the user task exactly
  2. What's the worst-case failure consequence
  3. Why isn't a simpler workflow sufficient
  4. How will quality be monitored post-launch
  5. Which layer can you roll back to when issues arise

If these 5 questions can't be answered clearly, hold off on drawing complex architecture diagrams.


Practice

Take your most-wanted AI feature. First determine which category it resembles:

  1. Knowledge-type problem
  2. Multi-step execution problem
  3. Both
  4. Actually just regular automation

Get the classification right, and the system design direction usually won't be too far off.

End the Discussion with an Architecture Decision Record

Do not use a complex diagram as a substitute for a product decision. A minimum ADR states:

FieldQuestion to answer
User taskFind an answer, generate content, or execute several actions?
Simplest optionCan it work without RAG or Agent?
ChoiceOrdinary automation, RAG, Agent, or a layered combination
EvidenceWhich evaluation shows the current option is insufficient?
PermissionWhat may the Agent read or write, and when must it confirm?
FailureHow does it degrade on missing sources, tool failure, or denied access?
Cost and latencyBudget and wait limit per successful task
RollbackWhich stable version can be restored?

Completion Criteria

  • A simpler alternative and the evidence for rejecting it are recorded
  • RAG has retrieval-failure samples; Agent has permissions and termination rules
  • Architecture choice connects to task metrics, not a model demo
  • Failure never silently executes or reports invented success

Chapter Deliverable

Complete an AI Architecture Decision Record, then continue to AI Ethics & Compliance and review data, permission, misuse, and human-takeover risks.

📚 Related resources

Common questions

Open a question to review the practical answer.

When should you use RAG and when should you use an Agent?

Missing knowledge → RAG (internal Q&A, support copilot, policy lookup, long-doc Q&A). Missing multi-step execution → Agent (research workflows, multi-tool calls, complex ops, semi-autonomous tasks). Both missing → RAG + Agent, layered. If it is a form or rule flow, you do not need AI system complexity at all—do not pick the most complex stack because it 'sounds advanced'.

When should you not rush to add RAG?

Four red flags: doc quality is poor (garbage in, garbage retrieved), there is no knowledge-refresh process (it goes stale fast), users do not actually want Q&A (you are optimizing the wrong problem), and the team has not defined source trust (retrieving is not the same as trusting). Most 'RAG works badly' complaints are knowledge-base governance failures, not model failures.

Why should a PM focus on Agent controllability rather than capability?

Agent autonomy amplifies risk. Ask: (1) can it call tools it should not (2) will it keep executing on a wrong premise (3) is each step observable (4) can it interrupt or escalate to a human on failure. If a single error is expensive, single-step quality is not solid yet, or users do not actually want autonomy, an Agent adds complexity without adding value.

Why must RAG and Agent be layered when used together?

Split it into knowledge layer → retrieval layer → decision layer → action layer, so you can tell whether retrieval was wrong, judgment was wrong, or tool execution was wrong. The biggest problem with a single 'do-everything Agent' is not performance—it is that when things break, nobody can pinpoint which layer broke.

What 5 questions must a team answer before evaluating a RAG or Agent design?

(1) What is the user's task, exactly? (2) What is the worst-case failure cost? (3) Why is a simpler workflow not enough? (4) How will quality be monitored post-launch? (5) Which layer can we roll back to when it breaks? If those five are foggy, hold off on the architecture diagram—technical complexity is not product progress.