Architecture

This document describes the architecture of Routir.

Overview

Routir is built with a modular architecture consisting of several key components:

• Models: Search and reranking models (PLAID-X, LSR, Qwen3, mT5, etc.)

• Processors: Request batching, caching, and content processing

• Pipeline: Composable search pipelines with multiple stages

• Config: JSON-based configuration system

Components

Models

Models implement the core search and scoring functionality. All models inherit from a base abstract class and implement the required interface methods.

Processors

Processors handle request batching, caching, and content transformation. They provide asynchronous and batch processing capabilities for improved performance.

Pipeline

The pipeline system allows composing multiple search stages together, enabling complex workflows like hybrid search, reranking, and score fusion.

Configuration

Services are configured using JSON files that specify models, processors, caching settings, and pipeline configurations.