What Are the Prompt Compression Modes in OmniRoute? A Complete Guide

OmniRoute supports seven distinct prompt compression modes—off, lite, standard, aggressive, ultra, rtk, and stacked—that let you trade latency for token economy using rule-based engines, SLM rewriting, or composable pipelines.

The diegosouzapw/OmniRoute repository implements a flexible compression framework that allows developers to select from predefined strategies or build custom pipelines. These prompt compression modes are defined in the core configuration layer and can be applied globally or per-request to optimize token usage.

The Seven Prompt Compression Modes Explained

In src/lib/db/compression.ts, the COMPRESSION_MODES constant enumerates the available strategies as a TypeScript set containing seven distinct options:

• off — Disables compression entirely; prompts are transmitted unchanged to the upstream model. This mode bypasses all transformation engines.
• lite — Runs a lightweight engine that performs fast, cheap transformations such as whitespace collapsing and system-prompt deduplication. Ideal for low-latency scenarios.
• standard — Activates the default "caveman" engine, a rule-based semantic condenser that applies balanced reductions including system-prompt deduplication and role compression.
• aggressive — Applies a heavier compressor that enforces summarization thresholds and stronger tool-output reductions. Configuration is controlled by the aggressive config block.
• ultra — Leverages a Small-Language-Model (SLM) to rewrite the prompt, aiming for maximal token savings while preserving semantic meaning. Managed via the ultra config block.
• rtk — Executes a rule-based "RTK" engine that filters and deduplicates tool results according to configurable business rules.
• stacked — Enables a composable pipeline that chains multiple engines (such as lite, caveman, aggressive, ultra, and rtk) in a user-defined order defined by the stackedPipeline setting.

How Modes Map to Compression Engines

According to the OmniRoute source code, the framework separates the user-facing mode from the concrete engine implementation using two key mappings in src/lib/db/compression.ts.

The SINGLE_MODE_ENGINE map resolves single-mode selections to their underlying engine IDs:

const SINGLE_MODE_ENGINE: Partial<Record<CompressionMode, string>> = {
  lite: "lite",
  standard: "caveman",
  aggressive: "aggressive",
  ultra: "ultra",
  rtk: "rtk",
};

When using stacked mode, the pipeline references engines listed in STACKED_PIPELINE_ENGINE_IDS, which includes:

const STACKED_PIPELINE_ENGINE_IDS = new Set([
  "lite",
  "caveman",
  "aggressive",
  "ultra",
  "rtk",
  "headroom",
  "session-dedup",
  "ccr",
  "llmlingua",
]);

Configuring Prompt Compression Modes

You can interact with these modes programmatically using the settings API exposed in src/lib/db/compression.ts.

Reading the Current Mode

import { getCompressionSettings } from "@/lib/db/compression";

async function showMode() {
  const cfg = await getCompressionSettings();
  console.log("Current mode:", cfg.defaultMode); // e.g. "standard"
}
showMode();

Switching to Ultra Compression

import { updateCompressionSettings } from "@/lib/db/compression";

async function enableUltra() {
  await updateCompressionSettings({ 
    defaultMode: "ultra", 
    ultra: { enabled: true } 
  });
}
enableUltra();

Building a Stacked Pipeline

Configure a custom sequence running lite → caveman → rtk:

import { updateCompressionSettings } from "@/lib/db/compression";

async function setStackedPipeline() {
  await updateCompressionSettings({
    defaultMode: "stacked",
    stackedPipeline: [
      { engine: "lite" },
      { engine: "caveman", intensity: "full" },
      { engine: "rtk", intensity: "standard" },
    ],
  });
}
setStackedPipeline();

Verifying Pipeline Configuration

import { getCompressionSettings } from "@/lib/db/compression";

async function logPipeline() {
  const cfg = await getCompressionSettings();
  console.log(cfg.stackedPipeline);
}
logPipeline();

Summary

• OmniRoute provides seven prompt compression modes: off, lite, standard, aggressive, ultra, rtk, and stacked.
• The COMPRESSION_MODES constant in src/lib/db/compression.ts defines the valid mode strings.
• Single modes map to engines via SINGLE_MODE_ENGINE, where standard maps to the caveman engine.
• Stacked mode allows composing engines from the STACKED_PIPELINE_ENGINE_IDS set, including specialized options like headroom, session-dedup, and llmlingua.
• Use getCompressionSettings() and updateCompressionSettings() to read and modify configurations programmatically.

Frequently Asked Questions

What is the difference between standard and ultra prompt compression modes?

Standard mode uses the caveman engine, a rule-based semantic condenser that applies balanced reductions like role compression and deduplication. Ultra mode instead invokes a Small-Language-Model (SLM) to rewrite the entire prompt, achieving higher token savings at the cost of increased latency and compute.

Can I combine multiple compression engines in OmniRoute?

Yes. The stacked mode allows you to compose a custom pipeline using the stackedPipeline configuration array. You can chain any engines listed in STACKED_PIPELINE_ENGINE_IDS, such as lite, caveman, and rtk, applying them sequentially in your specified order.

Where are compression settings stored in OmniRoute?

Compression configurations are persisted through the database layer defined in src/lib/db/compression.ts and src/lib/db/compressionCombos.ts. The getCompressionSettings() and updateCompressionSettings() functions handle retrieval and persistence, typically backed by the application's configured database.

How do I disable prompt compression entirely?

Set the defaultMode to off using updateCompressionSettings({ defaultMode: "off" }). This mode bypasses all transformation engines and sends prompts to the upstream model exactly as constructed, consuming the maximum token allowance but eliminating compression latency.

{
  "mcpServers": {
    "instagit": {
      "command": "npx",
      "args": ["-y", "instagit@latest"]
    }
  }
}