Skip to content

README.md

Latest commit

 

History

History
173 lines (125 loc) · 6.45 KB

README.md

File metadata and controls

173 lines (125 loc) · 6.45 KB
Raw

@routerlab/cli

route — the terminal interface to routerlab's cost-quality routing engine.

The CLI lets you route prompts, inspect published Pareto frontiers, list candidate models, and run the frontier eval pipeline — all from a shell. It wraps @routerlab/core so every routing decision is grounded in the same atlas-calibrated empirical token economics that powers the library.

Install

Runtime: the published CLI runs on Node.js >=20. This repository uses Bun for local workspace orchestration and tests.

npm install --save-dev @routerlab/cli

Or run without installing:

npx --yes @routerlab/cli route --task=qa --quality-bar=0.85 --input=prompt.txt

Or, inside the routerlab monorepo:

bun install
bun packages/cli/src/index.ts --help

Quick examples

# Route a prompt for QA at quality bar 0.85:
echo "What's the capital of France?" \
  | npx --yes @routerlab/cli route --task=qa --quality-bar=0.85

# Same, with a $0.005 hard budget and JSON output for piping into jq:
echo "Write a python function that ..." \
  | npx --yes @routerlab/cli route --task=codegen --quality-bar=0.80 --max-cost-usd=0.005 --json \
  | jq '.chosen.model.model'

# Read the prompt from a file instead of stdin:
npx --yes @routerlab/cli route --task=qa --quality-bar=0.85 --input=./prompt.txt

# Inspect the published Pareto frontier for codegen:
npx --yes @routerlab/cli frontier --task=codegen

# Same, as JSON:
npx --yes @routerlab/cli frontier --task=codegen --format=json

# List candidate models (optionally filtered by provider):
npx --yes @routerlab/cli models
npx --yes @routerlab/cli models --provider=anthropic
npx --yes @routerlab/cli models --json

# Run the frontier eval pipeline for a single task:
npx --yes @routerlab/cli eval frontier --task=qa --n=20

# Print versions:
npx --yes @routerlab/cli version

Subcommands

route route --task=<t> --quality-bar=<q> [flags]

Reads a prompt from --input <file> or stdin and prints the engine's decision: the chosen model, ordered fallbacks, and a skipped-with-reasons list. Default output is human-friendly; --json emits the raw RouteDecision for piping.

Flag Required Description
--task=<t> yes One of qa, codegen, summarization, classification, reasoning
--quality-bar=<q> yes Float in [0, 1]
--input=<path> no Read prompt from file (else stdin)
--max-cost-usd=<n> no Hard budget cap, USD
--max-latency-ms=<n> no Hard latency cap, milliseconds
--json no Emit raw RouteDecision JSON

Example output:

Decision: claude-sonnet-4-6 (anthropic)
  expected cost:    $0.001533
  expected quality: 0.900
  reasoning:        cheapest model meeting quality bar 0.850 for task "qa"; ...

Fallbacks (ordered by next-cheapest):
  1. claude-opus-4-7 (anthropic)   cost-rank 2 of 2 ...

Skipped (4):
  - claude-haiku-4-5: expected quality 0.800 for task "qa" is below quality bar 0.850
  - llama-3.3-70b: expected quality 0.800 for task "qa" is below quality bar 0.850
  - ...

route frontier --task=<t> [flags]

Reads eval/results/frontier.json and pretty-prints the Pareto frontier for the requested task. Useful for "what should I pick for codegen at quality bar 0.85?" Q&A from the terminal.

Flag Required Description
--task=<t> yes Task class
--format=table|json no Default table
--path=<p> no Override the frontier.json path (else uses repo default or ROUTERLAB_FRONTIER_PATH)

route models [flags]

Lists the engine's candidate model pool from packages/core/src/candidates.json.

Flag Description
--provider=<p> Filter to one of anthropic, openai, google, groq, together, hf, openrouter
--json JSON output

route eval frontier --task=<t> [--n=<n>]

Invokes the frontier builder at eval/frontier/runner.ts for a single task with n examples (default per the runner). Persists outcomes to eval/results/ and prints a JSON summary on stdout.

The runner is loaded dynamically; override its module path with ROUTERLAB_FRONTIER_RUNNER_MODULE if you've vendored it elsewhere.

route version

Prints @routerlab/cli and @routerlab/core versions.

Exit codes

Code Meaning
0 Success
1 No candidates pass the filters
2 Invalid input (bad flag, malformed value, missing arg)
3 Downstream error (calibration / runner / missing file)

These are stable. Script against them.

Programmatic use

The CLI is a thin wrapper around @routerlab/core. For library access:

import { route } from "@routerlab/core";

const decision = route({
  task: "qa",
  prompt: "What's the capital of France?",
  qualityBar: 0.85,
  maxCostUsd: 0.005,
});
console.log(decision.chosen.model.model);

Environment variables

Variable Effect
ROUTERLAB_FRONTIER_PATH Override the default frontier.json path
ROUTERLAB_FRONTIER_RUNNER_MODULE Override the dynamically-imported frontier runner module path
ROUTERLAB_ATLAS_RESULTS_PATH Atlas calibration file path (read by @routerlab/core/cost.ts)
ROUTERLAB_QUALITY_TABLE_PATH Quality table path (read by @routerlab/core/quality_predictor.ts)

License

Apache-2.0. See LICENSE.