11 Retrieval Systems 3 min read 559 words
Agentic RAG (Dec 2025)
Agentic RAG moves from a "Linear Pipeline" to a "Reasoning Loop." Instead of simply retrieving once, an agent decides when and what to retrieve to resolve a query.
Linear vs. Agentic RAG
| Model | Linear RAG | Agentic RAG |
|---|---|---|
| Structure | Predetermined sequence | Dynamic loop |
| Self-Correction | None | High (Can re-retrieve) |
| Query Complexity | Simple (1-step) | Hard (Multi-step) |
| Latency | Low (Fixed) | Variable (Multiple turns) |
2025 Principle: Use Agentic RAG when the query requires "Synthesized Proof" rather than just a "Document Match."
Self-RAG (Self-Reflection)
popularized in 2024/2025, Self-RAG uses "Critic Tokens" to evaluate its own work.
- Retrieve: Model pulls Top-K chunks.
- Evaluate: Is the info relevant? (CRITIC:
Relevant) - Generate: Is the answer supported? (CRITIC:
Supported) - Iterate: If the answer isn't supported, the model automatically triggers a broader search.
Corrective RAG (CRAG)
CRAG adds a "Reliability Layer" between retrieval and generation.
The Logic:
- If retrieval is Correct: Direct generation.
- If retrieval is Ambiguous: Use a Web-Search tool to supplement.
- If retrieval is Incorrect: Discard context and use external search or fallback logic.
Multi-Hop Reasoning Loops
For questions like "Who is the CEO of the company that acquired Figma?", the system must:
- Hop 1: Search for "Who acquired Figma?" (Result: Adobe).
- Hop 2: Search for "CEO of Adobe" (Result: Shantanu Narayen).
Agentic Pattern: The agent maintains a "State Object" and updates its "Sub-goal" after every retrieval until the chain is complete.
Agentic Filtering and Plan Revision
In late 2025, agents use Sub-Step Plans.
- Instead of one big retrieval, the agent writes a plan: "First I will check our internal database for X, then I will look at the public API for Y."
- Revised planning: If Step 1 fails, the agent rewrites Step 2.
Interview Questions
Q: What is the "Reasoning-Retrieval Balance" in Agentic RAG?
Strong answer: Every "Reasoning turn" in an agentic loop adds token cost and user latency. The goal of a production engineer is to find the "Retrieval Threshold." We use Token-Budgeting where we allow the agent only 3-5 "turns" before forcing a final answer. We also use Speculative Retrieval—where the agent predicts the next 2 steps it will take and retrieves for both simultaneously to reduce round-trip latency.
Q: Why does Agentic RAG often lead to higher quality but lower "Reliability" (Determinism)?
Strong answer: Agentic RAG is non-deterministic because the model is "Deciding" its path at every step. A small change in the user query might cause the agent to pick a different tool or search strategy, leading to a different answer format. In 2025, we mitigate this by using Constrained Agent Frameworks (like LangGraph or DSPy) where the "Graph of possible paths" is strictly defined, even if the choice between those paths is stochastic.
References
- Asai et al. "Self-RAG: Learning to Retrieve, Generate, and Critique" (2024/2025)
- Yan et al. "Corrective Retrieval Augmented Generation (CRAG)" (2024)
- LangChain. "Agentic RAG with LangGraph" (2025)
Key takeaways
01
Agentic RAG trades determinism for depth
A dynamic loop can re-retrieve and self-correct on multi-step questions, but the same query may take a different path each run, so latency and answer shape both vary.
02
Critic tokens close the loop
Self-RAG scores its own retrieval as relevant and its own answer as supported, and automatically triggers a broader search when the support check fails.
03
CRAG routes on retrieval quality
Correct retrieval generates directly, ambiguous retrieval is supplemented by web search, and incorrect retrieval has its context discarded rather than reasoned over.
04
Cap the turns before you ship
Token budgeting allows only three to five reasoning turns before a final answer is forced, and a constrained framework fixes the graph of paths even where the choice stays stochastic.