@paroicms/site-generator-plugin/backend/dist/generator/fake-content-generator/invoke-generate-fake-content.js

ParoiCMS Site Generator Plugin

import { languageLabelIn } from "@paroicms/public-anywhere-lib";
import { batchInvokeMinistral } from "../lib/calling-llm-mistral.js";
import { readPromptFile } from "../lib/create-prompt.js";
import { debugBatchLlmOutputs } from "../lib/debug-utils.js";
import { parseLlmResponseAsList } from "../lib/parse-llm-response.js";
import { buildPromptTemplate } from "../lib/prompt-template.js";
// Load and create prompt templates
const singlePromptTemplate = buildPromptTemplate(await readPromptFile("generate-fake-content-single.md"));
const multipleDocumentsPromptTemplate = buildPromptTemplate(await readPromptFile("generate-fake-content-multiple-documents.md"));
const multiplePartsPromptTemplate = buildPromptTemplate(await readPromptFile("generate-fake-content-multiple-parts.md"));
const multipleTermsPromptTemplate = buildPromptTemplate(await readPromptFile("generate-fake-content-multiple-terms.md"));
export async function invokeGenerateFakeContent(ctx, input, outputTags, options) {
    const { language, typeKind } = input;
    const single = input.count === 1;
    const llmTaskName = `fake-content-${options.llmTaskName}${single ? "" : `-${input.count}`}`;
    const tagAndDescriptions = outputTags
        .map(({ tagName, tagDescription }) => `<${tagName}>${tagDescription}</${tagName}>`)
        .join("\n");
    // Call LLM in batch mode
    const maxBatchSize = 15;
    const batchInputs = [];
    let startIndex = 0;
    while (startIndex < input.count) {
        const nextIndex = startIndex + maxBatchSize;
        const end = Math.min(nextIndex, input.count);
        const llmInput = {
            tagAndDescriptions,
            typeLabel: input.typeLabel,
            typeDescription: input.typeDescription,
            siteTheme: input.siteTheme,
            language: languageLabelIn(language, "en"),
        };
        if (typeKind === "part") {
            llmInput.documentDescription = input.documentDescription;
        }
        if (!single) {
            llmInput.count = `${end - startIndex}`;
        }
        batchInputs.push(llmInput);
        startIndex = nextIndex;
    }
    const debug = await debugBatchLlmOutputs(ctx, llmTaskName, ctx.mistralModelName, undefined, batchInputs);
    let llmOutput = debug.stored;
    if (!llmOutput) {
        // Select the appropriate prompt template
        const promptTemplate = single
            ? singlePromptTemplate
            : typeKind === "document"
                ? input.isTaxonomyTerm
                    ? multipleTermsPromptTemplate
                    : multipleDocumentsPromptTemplate
                : multiplePartsPromptTemplate;
        // Process batch inputs
        const messages = batchInputs.map(promptTemplate);
        const results = await batchInvokeMinistral(ctx, messages, {
            llmTaskName,
            temperature: 0.1,
            maxTokens: 50_000,
            timeoutMs: 60_000,
        });
        llmOutput = await debug.getMessageContents(results);
    }
    const results = [];
    for (const llmMessageContent of llmOutput.outputs) {
        const list = parseLlmResponseAsList(llmMessageContent, outputTags, {
            ...options,
            titleFromMarkdownHeading: outputTags.some((tag) => tag.key === "title")
                ? "title"
                : undefined,
        });
        results.push(...list.map((fields) => Object.fromEntries(Object.entries(fields).map(([fieldName, value]) => [fieldName, { [language]: value }]))));
    }
    return { contents: results, llmReport: llmOutput.llmReport };
}