Two-Tier Search Architecture
Definition (confidence: high)
A two-tier search architecture separates the planning (search, reasoning, research strategy) from the answer (writing, formatting, final output). The planning system issues queries, navigates sources, and explains its thought process. The answer system receives the search results and writes the final response, unaware of the planning tier’s internal reasoning.
Key Features (confidence: high)
- Separation of concerns: The planner optimizes for thorough research; the writer optimizes for clear, formatted output
- Invisible planning: The end user never sees the planning tier’s work — only the final answer
- Specialized prompts: Each tier has its own system prompt optimized for its role
- Deterministic handoff: The planner’s search results are the only data passed to the answer tier
Example: Perplexity AI
The perplexity-system-prompt from perplexity-ai implements this pattern:
- Planning tier: Has
[planning_rules]— defines how to determine query type, break down complex queries, assess sources, and explain reasoning - Answer tier: Has
[goal],[format_rules],[restrictions],[output]— focused entirely on writing a correct, high-quality final answer
Implications
This architecture reduces cognitive load on each tier — the planner doesn’t worry about formatting, and the writer doesn’t worry about search strategy. It also makes the system more auditable: the planner’s reasoning can be logged and reviewed without affecting the user-facing answer.
Open Questions
- Does this architecture leak latency (two LLM calls instead of one)?
- How does error recovery work if the planner fails (e.g., no useful search results)?
- Could this be extended to three tiers (plan → research → write)?