Ratio
System Architecture
Back

Interactive Pipeline

The core execution flow of the Ratio agent. Click any node in the graph below to inspect its operational logic and data transformations.

LOADING PIPELINE...

Technical Specifications

A complete breakdown of the models, infrastructure, and parameters powering the production deployment. Built for scale, low latency, and deterministic reasoning.

Retrieval Engines

Dense Vector ModelBAAI/bge-small-en-v1.5
Vector Dimensions384d
Sparse IndexBM25 (Lexical)
Reranking ModelBAAI/bge-reranker-base
Retrieval StrategyHybrid (Alpha=0.75) + Cross-Encoder

Agentic Reasoning

Base LLMIFM/K2-Horizon-375B-A23B
Inference ModeStreaming Server-Sent Events (SSE)
Reasoning LoopThought-Action-Observation (ReAct)
Context Window128,000 tokens
System PromptContextual Legal Analysis Constraints

Data Infrastructure

Vector DatabaseLanceDB (Memory-mapped columnar)
Relational DatabaseSQLite3 (Metadata & Lineage)
Total Corpus Size580,939 processed chunks
Data Sharding12 parallel shards
Chunking StrategyRecursive Character (1024 / 200)

Compute & Deployment

Backend APIFastAPI + Uvicorn
Runtime EnvironmentGoogle Cloud Run (Serverless)
Frontend FrameworkNext.js 14 App Router
Edge DeploymentVercel Edge Network
ContainerizationDocker (Multi-stage build)