A classic RAG receives a question, retrieves passages, and then generates an answer. This chain is fast, understandable, and suitable for many documentary uses. An agentic RAG adds a loop: the model analyzes the question, chooses a source or tool, examines the result, reformulates, and continues until a stopping criterion is met.
This autonomy can solve multi-step questions. It also increases the number of calls, variability, and attack surface. The goal is not to make all RAG agentic, but to reserve planning for tasks that need it.
The classic RAG as a baseline
A simple architecture includes:
- normalization of the question;2. full-text, vector, or hybrid search;3. rights filtering;4. possible reranking;5. building the context;6. generation with quotes;7. format control.
It is testable step by step. Latency and cost are predictable. A large part of the problems comes from the quality of the corpus, the segmentation, or the search, not from the absence of an agent.
The limits of a single research
A single request becomes insufficient when the question asks:
- compare several entities;- link events over time;- consult different sources;- calculate from results;- resolve an ambiguity;- verify a statement;- link research and business tool.
A question like 'which clauses changed between the two versions and what impact on the open files?' requires several steps and sources.
Before the agent: the intermediate strategies
Controlled reformulation
A rule or a pattern produces a few variants, then the results are merged. The number of calls remains bounded.
Decomposition according to a pattern
For a comparison, the system creates a sub-question for each entity and then gathers the answers. The structure is known and testable.
Routing
A classifier chooses the index, the language, or the tool. It does not plan freely.
Reranking and neighboring context
Sometimes, the relevant passage already exists in the candidates. Improving the ranking is enough.
These techniques offer some of the benefits with less variability.
What an agentic RAG adds
An agent maintains a search state, chooses a next action, and evaluates whether the information is sufficient. It can:
- ask for a clarification;- select a source;- generate a query;- call an API;- read a result;- to detect a gap;- verify a statement;- synthesize with origin.
Actions must remain within a bounded catalog and the plan is never an authorization.
The criteria that justify an agentic loop
Path variability
The questions require different steps that cannot be simply coded.
Multiple sources
Documents, database, authorized web, and business tools must be coordinated.
Verification possible
The system can control a sum, a quotation, consistency, or the presence of evidence.
Sufficient value
The gain on complex questions offsets the latency and cost.
Tolerance for uncertainty
The result is reviewed or used in a decision whose controls are adapted.
Assessment game
Multi-step cases allow measuring whether the loop actually brings an improvement.
Four levels of RAG, from single retrieval to iterative agentic search with verification.
Design the plan as an explicit state
The plan must not exist only in hidden text. The system keeps structured steps: goal, action, parameters, result, evidence, status, and next decision.
This representation allows:
- limitation of the number of steps;- resumption;- display to the user;- audit;- comparison;- interruption.
The private reasoning of the model is not necessary. It is enough to keep operational decisions and short justifications.
Choose bounded tools
A search tool receives a query, filters, and a maximum number. A calculator executes a defined operation. A business access respects the rights. The agent does not receive an unlimited shell, SQL, or browser.
Each tool returns structured data, a provenance, and errors. The results are considered unreliable until validated.
Define the stopping criteria
A loop without limit can continue to search, consume, and accumulate noise. The criteria include:
- maximum number of steps;- token budget or cost;- duration;- sufficient evidence;- absence of new information;- minimal confidence;- need clarification;- action prohibited.
When the criterion is not met, the system produces a partial response or an explicit refusal.
Check rather than self-convince
Asking the same model if it is satisfied does not constitute proof. Useful checks are external: existing references, calculation, diagram, business constraints, second research method, or human review.
A model can play a critical role, but its verdicts are calibrated on human cases and do not replace deterministic checks.
Manage multi-step quotes
Each final statement must be linked to the passages and results that support it. The system maintains the source at every step and prevents an intermediate summary from losing the references.
When a calculation uses multiple sources, display the entries, the versions, and the formula. Contradictions are flagged.
Security
The loop increases exposure to prompt injection: each document and tool response can attempt to influence the plan. The data remains separate from the policies. The tools and destinations are controlled by an independent layer.
The context of each step is minimal. Secrets are never accessible to the model. Web or code calls run in an isolated environment.
Latency and user experience
An agentic search can take several seconds or minutes. The interface shows the useful state: sources in progress, step, possibility to cancel, and partial result. It does not display pseudo-theatrical reasoning.
Long tasks can be asynchronous, with notification and resumption. A cache reuses stable sub-results when rights and freshness allow it.
Cost
The cost varies with the number of steps, requests, tokens, and rerankings. A budget per task and a routing model prevent costly loops.
Simple questions should take the direct path. A classifier or rules can reserve the agent for complex intentions.
Assess the contribution
Compare the agentic RAG to a strong baseline, not to a weak prototype. Measure:
- correct response rate;- coverage of evidence;- quality of refusals;- number of steps;- latency;- cost;- stability;- tool errors;- satisfaction.
Analysis by question type reveals where the agent is helpful and where it degrades the experience.
Deploy gradually
Start in observation mode: the agent proposes a plan without acting, then compare it to the current method. Then activate a few tools in reading mode, strict budgets, and a human review.
Failed cases feed the evaluation sequence. Limits are increased only if the gains are proven.
Simplicity as the default architecture
A well-designed classic RAG remains the best choice for a knowledge base and direct questions. Agentic search becomes relevant when the task truly requires variable and verifiable exploration.
Partitech can build a RAG baseline, design the search loops, secure the tools, and set up benchmarking evaluations. The goal is to add autonomy only where it improves useful quality.
Let's talk about your project
Assess the interest of an agentic RAG on your real cases with Partitech.Contact Partitech.