faithfulness.ts•3.93 kB
/**
* Copyright 2024 Google LLC
*
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
*/
import { z, type Genkit, type ModelArgument } from 'genkit';
import {
EvalStatusEnum,
type BaseEvalDataPoint,
type Score,
} from 'genkit/evaluator';
import path from 'path';
import { getDirName, loadPromptFile, renderText } from './helper.js';
const LongFormResponseSchema = z.object({ statements: z.array(z.string()) });
const NliResponseBaseSchema = z.object({
statement: z.string(),
reason: z.string(),
verdict: z.boolean(),
});
type NliResponseBase = z.infer<typeof NliResponseBaseSchema>;
const NliResponseSchema = z.object({
responses: z.array(NliResponseBaseSchema),
});
/**
*
*/
export async function faithfulnessScore<
CustomModelOptions extends z.ZodTypeAny,
>(
ai: Genkit,
judgeLlm: ModelArgument<CustomModelOptions>,
dataPoint: BaseEvalDataPoint,
judgeConfig?: CustomModelOptions
): Promise<Score> {
try {
if (!dataPoint.input) {
throw new Error('Input was not provided');
}
if (!dataPoint.output) {
throw new Error('Output was not provided');
}
if (!dataPoint.context?.length) {
throw new Error('Context was not provided');
}
const input =
typeof dataPoint.input === 'string'
? dataPoint.input
: JSON.stringify(dataPoint.input);
const output =
typeof dataPoint.output === 'string'
? dataPoint.output
: JSON.stringify(dataPoint.output);
const context = dataPoint.context.map((i) => JSON.stringify(i));
const longFormPrompt = await loadPromptFile(
path.resolve(getDirName(), '../../prompts/faithfulness_long_form.prompt')
);
const longFormResponse = await ai.generate({
model: judgeLlm,
config: judgeConfig,
prompt: await renderText(longFormPrompt, {
question: input,
answer: output,
}),
output: {
schema: LongFormResponseSchema,
},
});
const parsedLongFormResponse = longFormResponse.output;
const statements = parsedLongFormResponse?.statements ?? [];
if (statements.length === 0) {
throw new Error('No statements returned');
}
const allStatements = statements.map((s) => `statement: ${s}`).join('\n');
const allContext = context.join('\n');
const nliPrompt = await loadPromptFile(
path.resolve(getDirName(), '../../prompts/faithfulness_nli.prompt')
);
const response = await ai.generate({
model: judgeLlm,
prompt: await renderText(nliPrompt, {
context: allContext,
statements: allStatements,
}),
output: {
schema: NliResponseSchema,
},
});
const parsedResponse = response.output;
return nliResponseToScore(parsedResponse?.responses ?? []);
} catch (err) {
console.debug(
`Genkit faithfulness evaluation failed with error ${err} for sample ${JSON.stringify(
dataPoint
)}`
);
throw err;
}
}
function nliResponseToScore(input: NliResponseBase[] | null): Score {
if (!input) {
throw new Error(`Evaluator response empty`);
}
const faithfulStatements = input.reduce((total, resp) => {
return total + (resp.verdict ? 1 : 0);
}, 0);
const score = faithfulStatements / input.length;
return {
score,
details: { reasoning: input.map((r) => r.reason).join('; ') },
status: score > 0.5 ? EvalStatusEnum.PASS : EvalStatusEnum.FAIL,
};
}