Module search_plan 

Quantization-aware Search Plan

This module provides a formal runtime plan for vector search that separates policy (what to optimize for) from mechanism (how to execute).

Architecture

SearchRequest + SLA → Planner → SearchPlan → Executor → Results
                         ↑
                   Cost Model + Statistics

Policy vs Mechanism

Policy (what to optimize):

• Target recall@k (e.g., 0.95)
• Latency budget (e.g., 5ms p99)
• Token/compute budget

Mechanism (how to execute):

• BPS coarse scan parameters
• PQ scoring parameters
• Rerank depth and method
• ef_search value
• Filter evaluation order

Cost Model

The planner uses measured per-stage costs:

• cost_bps(N, D) = N × D × c_bps
• cost_pq(N, D, M) = N × M × c_pq
• cost_rerank(N, D) = N × D × c_f32

Optimization

Minimize expected latency subject to:

• recall@k ≥ target_recall
• total_cost ≤ budget

Uses bandit-like adaptation based on recent query statistics.

Structs

CostModel

Cost model parameters (calibrated per hardware).

DatasetStats

Statistics about the dataset for planning.

PipelineStage

A single stage in the search pipeline.

PlanExecutor

Plan executor that runs a search plan.

SearchPlan

The search plan: a complete specification for executing a search.

SearchPlanner

Search planner that generates optimal plans.

SearchSLA

Service Level Agreement for search.

StageCosts

Per-stage cost measurements.

Enums

OptimizationMode

Optimization mode for the planner.

PlanError

Plan validation errors.

StageQuantLevel

Quantization level for a pipeline stage.