Guardrails

Evaluation and guardrails are 2 sides of the same coin

Guardrails are just evaluators that run online. If you could detect a hallucination fast enough, why not catch that in the application's runtime and prevent the issue from surfacing to the user?

Because AutoEval models are fast (< 300ms on CPU and < 10ms on GPU) and cheap to operate (1/1000th the size of GPT-4), you can run them on every response as a guardrail without compromising on user experience.

info

Coming soon: Trust and safety features that allow you to provision guardrail endpoints and set appropriate thresholds for pass/fail.

Usage Guide

Use the API to compute metrics in the application runtime:

python

from lastmile.lib.auto_eval import AutoEval, Metric
import pandas as pd

def guard(input, output, context, metric: Metric, threshold=0.5) -> bool:
  client = AutoEval(api_token="api_token_if_LASTMILE_API_TOKEN_not_set")

  result = client.evaluate_data(
      data=pd.DataFrame({
          "input": [input],
          "output": [output],
          "ground_truth": [context]
      }),
      metrics=[metric]
  )

  score = result[f"{metric.name}_score"][0]
  return score >= threshold

# Define a faithfulness guardrail:
guard(
  input="Where did the author grow up?",
  output="France",
  context="England",
  metric=Metric(name="Faithfulness")
)

node.js

import { AutoEval, Metric } from "lastmile/lib/auto_eval";

async function guard(
  input: string,
  output: string,
  context: string,
  metric: Metric,
  threshold: number = 0.5
): Promise<boolean> {
  const client = new AutoEval({
    apiKey: "api_token_if_LASTMILE_API_TOKEN_not_set",
  });

  // Evaluate the data
  const result = await client.evaluateData(
    /*data*/ [
      {
        "input": input,
        "output": output,
        "ground_truth": context
      }
    ], 
    metric
  );

  // Extract the score
  const scoreColumnName = `${metric.name}_score`;
  const score = result[0][scoreColumnName];

  // Return whether the score meets the threshold
  return score >= threshold;
}

const faithfulnessMetric: Metric = { name: "Faithfulness" };

const isFaithful = await guard(
  /*input*/   "Where did the author grow up?",
  /*output*/  "France",
  /*context*/ "England",
  faithfulnessMetric
);

console.log(`Is the response faithful? ${isFaithful}`);

Fine-tune a custom guardrail

You can build your own input/output guardrail models using the fine-tuning service. For example, you can define a custom input guardrail that determines if the input is relevant to the purpose of the application, and denies irrelevant requests.

Follow this guide for a detailed walkthrough:

Real-time guardrails>

Build real-time guardrails in a RAG application using fine-tuned alBERTa 🍁 models.