Proactive planning structures an information need before retrieval begins. Instead of sending a complex question directly to one vector search, a high-level searcher decomposes it into atomic sub-queries, then coordinates retrieval and later verification. section 11.0
A plan-and-refine retrieval architecture typically:
- Decomposes a multi-hop question into atomic sub-queries. section 11.0
- Routes each sub-query to an appropriate search modality, such as sparse keyword search for entities or dense search for themes. section 11.0
- Synthesizes the retrieved results, checks whether they answer the complete question, and creates supplementary plans when gaps remain. section 11.0
Comparing policies across documents
A question such as “Compare Policy A and Policy B” is treated as a multi-document research task rather than one undifferentiated search. The planner creates atomic searches, executes them through low-level searchers, and verifies that the combined evidence supports the comparison. section 11.4
Table 11.1 Strategic selection matrix
|
Feature
|
CRAG (Corrective)
|
LevelRAG (Plan-and-Refine)
|
|
Primary Mechanism
|
Evaluation & Fallback
|
Decomposition & Hierarchical Search
|
|
Best Use Case
|
Fact-Checking, Real-time Validation
|
Complex Research, Multi-document Synthesis
|
|
Key Innovation
|
Decompose-then-Recompose
|
High-Level Logic Planning
|
|
Compliance Role
|
Screening ("Is X a PEP?")
|
Advisory ("Compare Policy A vs B")
|
|
Latency Profile
|
Low (Internal) / High (Web Fallback)
|
High (Iterative Planning)
|
|
Architecture
|
Component Add-on
|
Full System Architecture
|
Use the strategic selection matrix to compare proactive planning with reactive correction across mechanism, use case, latency, compliance role, and architecture. section 11.4
