Files
mintel.me/scripts/ai-estimate.ts
Marc Mintel c5ad6108ee
Some checks failed
Build & Deploy Mintel Blog / build-and-deploy (push) Failing after 2m28s
feat: Rename "Betrieb & Hosting" to "Sorglos-Paket" and refine PDF pricing module layout and text.
2026-02-05 01:53:12 +01:00

858 lines
44 KiB
TypeScript

import { CheerioCrawler, RequestQueue } from 'crawlee';
import * as path from 'node:path';
import * as fs from 'node:fs/promises';
import { existsSync } from 'node:fs';
import { URL } from 'node:url';
import { execSync } from 'node:child_process';
import axios from 'axios';
import { FileCacheAdapter } from '../src/utils/cache/file-adapter.js';
import { initialState, PRICING } from '../src/logic/pricing/constants.js';
import { calculateTotals } from '../src/logic/pricing/calculator.js';
async function main() {
const OPENROUTER_KEY = process.env.OPENROUTER_KEY;
if (!OPENROUTER_KEY) {
console.error('❌ Error: OPENROUTER_KEY not found in environment.');
process.exit(1);
}
let briefing = '';
let targetUrl: string | null = null;
let comments: string | null = null;
let budget: string | null = null;
let cacheKey: string | null = null;
let jsonStatePath: string | null = null;
const isEstimation = process.argv.includes('--estimation') || process.argv.includes('-E');
const args = process.argv.slice(2);
for (let i = 0; i < args.length; i++) {
const arg = args[i];
if (arg === '--url') {
targetUrl = args[++i];
} else if (arg === '--comments' || arg === '--notes') {
comments = args[++i];
} else if (arg === '--budget') {
budget = args[++i];
} else if (arg === '--cache-key') {
cacheKey = args[++i];
} else if (arg === '--json') {
jsonStatePath = args[++i];
} else if (arg === '--estimation' || arg === '-E') {
// Handled above
} else if (!arg.startsWith('--')) {
briefing = arg;
}
}
if (briefing && briefing.startsWith('@')) {
const rawPath = briefing.substring(1);
const filePath = rawPath.startsWith('/') ? rawPath : path.resolve(process.cwd(), rawPath);
briefing = await fs.readFile(filePath, 'utf8');
}
// Discovery ONLY if not provided
if (!targetUrl && briefing) {
const urlMatch = briefing.match(/https?:\/\/[^\s]+/);
if (urlMatch) {
targetUrl = urlMatch[0];
console.log(`🔗 Discovered URL in briefing: ${targetUrl}`);
}
}
if (!briefing && !targetUrl && !comments && !jsonStatePath) {
console.error('❌ Usage: npm run ai-estimate -- "Briefing text" [--url https://example.com] [--comments "Manual notes"]');
console.error(' Or: npm run ai-estimate -- @briefing.txt [--url https://example.com]');
console.error(' Or: npm run ai-estimate -- --json path/to/state.json');
process.exit(1);
}
const clearCache = process.argv.includes('--clear-cache');
if (clearCache) {
console.log('🧹 Clearing cache...');
const cacheFiles = await fs.readdir(path.join(process.cwd(), '.cache'));
for (const file of cacheFiles) {
if (file.startsWith('ai_est_')) {
await fs.unlink(path.join(process.cwd(), '.cache', file));
}
}
}
const cache = new FileCacheAdapter({ prefix: 'ai_est_' });
const finalCacheKey = cacheKey || `${briefing}_${targetUrl}_${comments}_${budget}`;
// 1. Crawl if URL provided
let crawlContext = '';
if (targetUrl) {
console.log(`🔍 Crawling ${targetUrl} for context...`);
const cachedCrawl = await cache.get<string>(`crawl_${targetUrl}`);
if (cachedCrawl && !clearCache) {
console.log('📦 Using cached crawl results.');
crawlContext = cachedCrawl;
} else {
crawlContext = await performCrawl(targetUrl);
await cache.set(`crawl_${targetUrl}`, crawlContext, 86400); // 24h cache
}
}
// 2. Distill Crawl Context (Context Filtering)
let distilledCrawl = '';
if (crawlContext) {
const cachedDistilled = await cache.get<string>(`distilled_${targetUrl}`);
if (cachedDistilled && !clearCache) {
distilledCrawl = cachedDistilled;
} else {
distilledCrawl = await distillCrawlContext(crawlContext, OPENROUTER_KEY);
await cache.set(`distilled_${targetUrl}`, distilledCrawl, 86400);
}
} else if (targetUrl) {
distilledCrawl = `WARNING: The crawl of ${targetUrl} failed (ENOTFOUND or timeout).
The AI must NOT hallucinate details about the current website.
Focus ONLY on the BRIEFING provided. If details are missing, mark them as 'unknown'.`;
console.warn('⚠️ Crawl failed. AI will be notified to avoid hallucinations.');
}
// 3. AI Prompting
console.log('🤖 Consultating Gemini 3 Flash...');
const cachedAi = !clearCache ? await cache.get<any>(finalCacheKey) : null;
let formState: any;
let usage: { prompt: number, completion: number, cost: number } = { prompt: 0, completion: 0, cost: 0 };
// Load Context Documents
const principles = await fs.readFile(path.resolve(process.cwd(), 'docs/PRINCIPLES.md'), 'utf8');
const techStandards = await fs.readFile(path.resolve(process.cwd(), 'docs/TECH.md'), 'utf8');
const tone = await fs.readFile(path.resolve(process.cwd(), 'docs/TONE.md'), 'utf8');
if (jsonStatePath) {
console.log(`📂 Loading state from JSON: ${jsonStatePath}`);
const rawJson = await fs.readFile(path.resolve(process.cwd(), jsonStatePath), 'utf8');
formState = JSON.parse(rawJson);
} else if (cachedAi) {
console.log('📦 Using cached AI response.');
formState = cachedAi;
} else {
const result = await getAiEstimation(briefing, distilledCrawl, comments, budget, OPENROUTER_KEY, principles, techStandards, tone);
formState = result.state;
usage = result.usage;
await cache.set(finalCacheKey, formState);
}
// 3. Save Data & Generate PDF
const timestamp = new Date().toISOString().replace(/[:.]/g, '-');
const jsonOutDir = path.resolve(process.cwd(), 'out/estimations/json');
if (!existsSync(jsonOutDir)) await fs.mkdir(jsonOutDir, { recursive: true });
const finalJsonPath = path.join(jsonOutDir, `${formState.companyName || 'unknown'}_${timestamp}.json`);
await fs.writeFile(finalJsonPath, JSON.stringify(formState, null, 2));
const tempJsonPath = path.resolve(process.cwd(), '.cache', `temp_state_${Date.now()}.json`);
await fs.writeFile(tempJsonPath, JSON.stringify(formState, null, 2));
console.log(`📦 Saved detailed state to: ${finalJsonPath}`);
console.log('📄 Generating PDF estimation...');
try {
const genArgs = isEstimation ? '--estimation' : '';
execSync(`npx tsx ./scripts/generate-estimate.ts --input ${tempJsonPath} ${genArgs}`, { stdio: 'inherit' });
} finally {
// await fs.unlink(tempJsonPath);
}
console.log('\n✨ AI Estimation Complete!');
if (usage.prompt > 0) {
console.log('--------------------------------------------------');
console.log('📊 ACCUMULATED API USAGE (SUM OF 6 PASSES)');
console.log(` Model: google/gemini-3-flash-preview`);
console.log(` Total Prompt: ${usage.prompt.toLocaleString()}`);
console.log(` Total Completion: ${usage.completion.toLocaleString()}`);
console.log(` Total Tokens: ${(usage.prompt + usage.completion).toLocaleString()}`);
console.log(` Total Cost (USD): $${usage.cost.toFixed(6)}`);
console.log('--------------------------------------------------\n');
}
}
async function distillCrawlContext(rawCrawl: string, apiKey: string): Promise<string> {
if (!rawCrawl || rawCrawl.trim().length === 0) return "Keine Crawl-Daten vorhanden.";
console.log(' ↳ Distilling Crawl Context (Noise Filtering)...');
const systemPrompt = `
You are a context distiller. Your goal is to strip away HTML noise, legal footers, and generic fluff from a website crawl.
Extract the "Company DNA" in 5-8 bullet points (GERMAN).
### FOCUS ON:
1. Core Business / Services.
2. Unique Selling Points (USPs).
3. Target Audience (if clear).
4. Tech Stack or industry-specific equipment mentioned.
5. Brand tone (e.g. "industrial", "friendly", "technical").
### OUTPUT:
Return ONLY the bullet points. No intro/outro.
`;
const resp = await axios.post('https://openrouter.ai/api/v1/chat/completions', {
model: 'google/gemini-3-flash-preview',
messages: [{ role: 'system', content: systemPrompt }, { role: 'user', content: `RAW_CRAWL:\n${rawCrawl.substring(0, 30000)}` }],
}, { headers: { 'Authorization': `Bearer ${apiKey}`, 'Content-Type': 'application/json' } });
return resp.data.choices[0].message.content;
}
async function performCrawl(url: string): Promise<string> {
const pages: { url: string, content: string, type: string }[] = [];
const origin = new URL(url).origin;
const crawler = new CheerioCrawler({
maxRequestsPerCrawl: 20,
async requestHandler({ $, request, enqueueLinks }) {
const title = $('title').text();
const urlObj = new URL(request.url);
const urlPath = urlObj.pathname.toLowerCase();
let type = 'other';
if (urlPath === '/' || urlPath === '') type = 'home';
else if (urlPath.includes('service') || urlPath.includes('leistung')) type = 'service';
else if (urlPath.includes('blog') || urlPath.includes('news') || urlPath.includes('aktuelles') || urlPath.includes('magazin')) type = 'blog';
else if (urlPath.includes('contact') || urlPath.includes('kontakt')) type = 'contact';
else if (urlPath.includes('job') || urlPath.includes('karriere') || urlPath.includes('career') || urlPath.includes('human-resources')) type = 'career';
else if (urlPath.includes('portfolio') || urlPath.includes('referenz') || urlPath.includes('projekt') || urlPath.includes('case-study')) type = 'portfolio';
else if (urlPath.includes('legal') || urlPath.includes('impressum') || urlPath.includes('datenschutz') || urlPath.includes('privacy')) type = 'legal';
const h1s = $('h1').map((_, el) => $(el).text()).get();
const navLinks = $('nav a').map((_, el) => $(el).text().trim()).get().filter(t => t.length > 0);
const bodyText = $('body').text().replace(/\s+/g, ' ').substring(0, 50000);
const html = $.html();
const hexColors = html.match(/#(?:[0-9a-fA-F]{3}){1,2}\b/g) || [];
const uniqueColors = Array.from(new Set(hexColors)).slice(0, 5);
pages.push({
url: request.url,
type,
content: `Title: ${title}\nType: ${type}\nHeadings: ${h1s.join(', ')}\nNav: ${navLinks.join(', ')}\nColors: ${uniqueColors.join(', ')}\nText: ${bodyText}`
});
await enqueueLinks({
limit: 15,
transformRequestFunction: (req) => {
const reqUrl = new URL(req.url);
if (reqUrl.origin !== origin) return false;
// Skip assets
if (reqUrl.pathname.match(/\.(pdf|zip|jpg|png|svg|webp)$/i)) return false;
return req;
}
});
},
});
await crawler.run([url]);
const typeCounts = pages.reduce((acc, p) => {
acc[p.type] = (acc[p.type] || 0) + 1;
return acc;
}, {} as Record<string, number>);
let summary = `\nCrawl Summary: Identified ${pages.length} pages total on ${origin}.\n`;
summary += Object.entries(typeCounts).map(([type, count]) => `- ${type}: ${count}`).join('\n') + '\n\n';
return summary + pages.map(p => `--- PAGE: ${p.url} ---\n${p.content}`).join('\n\n');
}
const cleanJson = (str: string) => {
// Remove markdown code blocks if present
let cleaned = str.replace(/```json\n?|```/g, '').trim();
// Remove potential control characters that break JSON.parse
// We keep \n \r \t for now as they might be escaped or need handling
cleaned = cleaned.replace(/[\u0000-\u0009\u000B\u000C\u000E-\u001F\u007F-\u009F]/g, " ");
// Specific fix for Gemini: raw newlines inside strings
// This is tricky. We'll try to escape newlines that are NOT followed by a quote and colon (property start)
// or a closing brace/bracket. This is heuristic.
// A better way is to replace all raw newlines that are preceded by a non-backslash with \n
// but only if they are inside double quotes.
// Simplest robust approach: Remove trailing commas and hope response_format does its job.
cleaned = cleaned.replace(/,\s*([\]}])/g, '$1');
return cleaned;
};
const getAiEstimation = async (briefing: string, distilledCrawl: string, comments: string | null, budget: string | null, apiKey: string, principles: string, techStandards: string, tone: string) => {
let usage = { prompt: 0, completion: 0, cost: 0 };
const addUsage = (data: any) => {
if (data?.usage) {
usage.prompt += data.usage.prompt_tokens || 0;
usage.completion += data.usage.completion_tokens || 0;
// OpenRouter provides 'cost' field in USD (as per documentation)
// If missing, we use a fallback calculation for transparency
if (data.usage.cost !== undefined) {
usage.cost += data.usage.cost;
} else {
// Fallback: Gemini 3 Flash Flash pricing (~$0.1 / 1M prompt, ~$0.4 / 1M completion)
usage.cost += (data.usage.prompt_tokens || 0) * (0.1 / 1000000) + (data.usage.completion_tokens || 0) * (0.4 / 1000000);
}
}
};
// 1. PASS 1: Fact Extraction (Briefing Sensor)
console.log(' ↳ Pass 1: Fact Extraction (Briefing Sensor)...');
const pass1SystemPrompt = `
You are a precision sensor. Analyze the BRIEFING and extract ONLY the raw facts.
Tone: Literal, non-interpretive.
Output language: GERMAN (Strict).
Output format: ROOT LEVEL JSON (No wrapper keys like '0' or 'data').
### MISSION:
Focus 100% on the BRIEFING text provided by the user. Use the DISTILLED_CRAWL only as background context for terms or company details. If there is a conflict, the BRIEFING is the absolute source of truth.
### PRICING REFERENCE (FOR CALCULATION):
- Base Project (Infrastructure + 12 Months Hosting): 5.440 € (MANDATORY START)
- Additional Pages: 600 € / stk
- System-Modules (Features): 1.500 € / stk
- Logic-Functions: 800 € / stk
- API Integrations: 800 € / stk
- CMS Setup: 1.500 € (optional)
- Visual Staging/Interactions: 1.500 € - 2.000 €
${budget ? `### BUDGET LOGIC (ULTRA-STRICT):
1. **Mental Calculation**: Start with 7.000 €. Add items based on the reference above.
2. **Hard Ceiling**: If total > ${budget}, you MUST discard lower priority items.
3. **Priority**: High-End Design and Core Pages > Features.
4. **Restriction**: For ${budget}, do NOT exceed 2 features and 4 extra pages.
5. THE TOTAL COST CALCULATED BY THESE RULES MUST BE <= ${budget}.
6. Do NOT mention the budget in any string fields.` : ''}
- ** features **: Items from the FEATURE_REFERENCE.
- ** ABSOLUTE CONSERVATIVE RULE **: Only use features if the briefing implies *dynamic complexity* (CMS, filtering, search, database).
- Simple keywords like 'Karriere', 'Referenzen', 'Messen' or lists of items MUST be treated as simple pages. Add them to 'otherPages' instead.
- If in doubt, categorizing as a PAGE is the mandatory default.
- ** otherPages **: Any specific pages mentioned (e.g. 'Historie', 'Team', 'Partner') that are not in the standard list. Use this for static lists of jobs or references too.
- ** companyName **: The full legal and brand name (e.g., "E-TIB GmbH"). Use signatures and crawl data.
- ** personName **: The name of the primary human contact (e.g., "Danny Joseph"). ** CRITICAL **: Check email signatures and "Mit freundlichen Grüßen" blocks.
- ** email **: The email address of the contact person if found in the briefing / signature.
- ** existingWebsite **: The primary URL mentioned in the briefing or signature (e.g., "www.e-tib.com").
- ** websiteTopic **: A short descriptor of the CORE BUSINESS (e.g., "Kabeltiefbau"). MAX 3 WORDS.
- ** isRelaunch **: Set to TRUE if the briefing mentions an existing website, a URL, or if the company is an established entity.
- ** CRITICAL LOGIC **: If a URL is mentioned, isRelaunch MUST be TRUE.
- For all textual values: USE GERMAN.
- ** multilang **: ONLY if the briefing mentions multiple target languages.
- ** maps **: If "Google Maps" or location maps are mentioned or implicit.
- ** CRITICAL **: Do NOT include "social" in apiSystems unless the user explicitly wants to SYNC / POST content.
### CATEGORY MAPPING(IDs ONLY):
- ** selectedPages **: [Home, About, Services, Contact, Landing, Legal]
- ** features **: [blog_news, products, jobs, refs, events]
- ** functions **: [search, filter, pdf, forms, members, calendar, multilang, chat]
- ** apiSystems **: [crm_erp, payment, marketing, ecommerce, maps, social, analytics]
- ** assets **: [existing_website, logo, styleguide, content_concept, media, icons, illustrations, fonts]
### OUTPUT FORMAT(Strict JSON - ROOT LEVEL):
{
"companyName": string,
"companyAddress": string,
"personName": string,
"email": string,
"existingWebsite": string,
"websiteTopic": string,
"isRelaunch": boolean,
"selectedPages": string[],
"features": string[],
"functions": string[],
"apiSystems": string[],
"assets": string[],
"deadline": string(GERMAN),
"targetAudience": "B2B" | "B2C" | "Internal" | string(GERMAN),
"expectedAdjustments": "low" | "medium" | "high" | string(GERMAN),
"employeeCount": "ca. 10+" | "ca. 50+" | "ca. 100+" | "ca. 250+" | "ca. 500+" | "ca. 1000+"
}
`;
const pass1UserPrompt = `BRIEFING(TRUTH SOURCE): \n${briefing} \n\nCOMMENTS: \n${comments} \n\nDISTILLED_CRAWL(CONTEXT ONLY): \n${distilledCrawl} `;
const p1Resp = await axios.post('https://openrouter.ai/api/v1/chat/completions', {
model: 'google/gemini-3-flash-preview',
messages: [{ role: 'system', content: pass1SystemPrompt }, { role: 'user', content: pass1UserPrompt }],
response_format: { type: 'json_object' }
}, { headers: { 'Authorization': `Bearer ${apiKey} `, 'Content-Type': 'application/json' } });
if (!p1Resp.data.choices?.[0]?.message?.content) {
console.error('❌ Pass 1 failed. Response:', JSON.stringify(p1Resp.data, null, 2));
throw new Error('Pass 1: No content in response');
}
const facts = JSON.parse(cleanJson(p1Resp.data.choices[0].message.content));
// 1.5. PASS 1.5: The Feature Auditor (Skeptical Review)
console.log(' ↳ Pass 1.5: The Feature Auditor (Skeptical Review)...');
const pass15SystemPrompt = `
You are a "Strict Cost Controller". Your mission is to prevent over-billing.
Review the extracted FEATURES and the BRIEFING.
### RULE OF THUMB:
- A "Feature" (1.500 €) is ONLY justified for complex, dynamic systems (logic, database, CMS-driven management, advanced filtering).
- Simple lists, information sections, or static descriptions (e.g., "Messen", "Team", "Historie", "Jobs" as mere text) are ALWAYS "Pages" (600 €).
- If the briefing doesn't explicitly mention "Management System", "Filterable Database", or "Client Login", it is likely a PAGE.
### MISSION:
Analyze each feature in the list. Decide if it should stay a "Feature" or be downgraded to the "otherPages" array.
### OUTPUT FORMAT:
Return only the corrected 'features' and 'otherPages' arrays.
{
"features": string[],
"otherPages": string[]
}
`;
const p15Resp = await axios.post('https://openrouter.ai/api/v1/chat/completions', {
model: 'google/gemini-3-flash-preview',
messages: [
{ role: 'system', content: pass15SystemPrompt },
{ role: 'user', content: `EXTRACTED_FEATURES: ${JSON.stringify(facts.features)} \nBRIEFING: \n${briefing}` }
],
response_format: { type: 'json_object' }
}, { headers: { 'Authorization': `Bearer ${apiKey}`, 'Content-Type': 'application/json' } });
addUsage(p15Resp.data);
const auditResult = JSON.parse(cleanJson(p15Resp.data.choices[0].message.content));
// Apply Audit: Downgrade features to otherPages
facts.features = auditResult.features || [];
facts.otherPages = Array.from(new Set([...(facts.otherPages || []), ...(auditResult.otherPages || [])]));
// 2. PASS 2: Feature Deep-Dive
console.log(' ↳ Pass 2: Feature Deep-Dive...');
const pass2SystemPrompt = `
You are a detail - oriented Solution Architect.
For EVERY item selected in Pass 1(pages, features, functions, apiSystems), write a specific justification and technical scope.
### RULES:
1. ** CONCRETE & SPECIFIC **: Do NOT say "Implementation of X". Say "Displaying X with Y filters".
2. ** JUSTIFICATION (CRITICAL) **: For every entry in 'featureDetails', explicitly explain WHY this is a complex system (1.500 €) and not just a static page (600 €). If it's just a list of items, it's a PAGE.
3. ** NO EFFECTS **: Do not mention "fade-ins", "animations" or "visual styling". Focus on FUNCTION.
4. ** ABSOLUTE RULE **: EVERYTHING MUST BE GERMAN.
5. ** TRANSPARENCY **: Explain exactly what the USER gets.
6. ** API NOTE **: For 'media' or 'video', explicitly state "Upload & Integration" (NO STREAMING).
### INPUT (from Pass 1):
${JSON.stringify(facts, null, 2)}
### OUTPUT FORMAT(Strict JSON):
{
"pageDetails": { "Home": string, ... },
"featureDetails": { "blog_news": string, ... },
"functionDetails": { "search": string, ... },
"apiDetails": { "crm_erp": string, ... }
}
`;
const p2Resp = await axios.post('https://openrouter.ai/api/v1/chat/completions', {
model: 'google/gemini-3-flash-preview',
messages: [{ role: 'system', content: pass2SystemPrompt }, { role: 'user', content: briefing }],
response_format: { type: 'json_object' }
}, { headers: { 'Authorization': `Bearer ${apiKey} `, 'Content-Type': 'application/json' } });
addUsage(p2Resp.data);
if (!p2Resp.data.choices?.[0]?.message?.content) {
console.error('❌ Pass 2 failed. Response:', JSON.stringify(p2Resp.data, null, 2));
throw new Error('Pass 2: No content in response');
}
const details = JSON.parse(cleanJson(p2Resp.data.choices[0].message.content));
// 3. PASS 3: Strategic Content (Bespoke Strategy)
console.log(' ↳ Pass 3: Strategic Content (Bespoke Strategy)...');
const pass3SystemPrompt = `
You are a high - end Digital Architect.Your goal is to make the CUSTOMER feel 100 % understood.
Analyze the BRIEFING and the EXISTING WEBSITE context.
### TONE & COMMUNICATION PRINCIPLES(STRICT):
${tone}
### OBJECTIVE:
3. ** briefingSummary **: Ein sachlicher, tiefgehender Überblick der Unternehmenslage.
- ** STIL **: Keine Ich - Form.Keine Marketing - Floskeln.Nutze präzise Fachbegriffe.Sei prägnant und effizient(ca. 70 % der vorherigen Länge).
- ** FORM **: EXAKT ZWEI ABSÄTZE.Insgesamt ca. 6 Sätze.
- ** INHALT **: Welcher technologische Sprung ist notwendig ? Was ist der Status Quo ? (Bezug zur URL / Briefing).
- ** ABSOLUTE REGEL **: Keine Halluzinationen über fehlende Präsenzen bei Relaunches.
- ** DATENSCHUTZ **: KEINERLEI namentliche Nennungen von Personen(z.B. "Danny Joseph") in diesen Texten.
4. ** designVision **: Ein abstraktes, strategisches Konzept.
- ** STIL **: Rein konzeptionell.Keine Umsetzungsschritte.Keinerlei "To-dos".Keine Ich - Form.Sei prägnant.
- ** FORM **: EXAKT ZWEI ABSÄTZE.Insgesamt ca. 4 Sätze.
- ** DATENSCHUTZ **: KEINERLEI namentliche Nennungen von Personen in diesen Texten.
- ** FOKUS **: Welche strategische Wirkung soll erzielt werden ? (Z.B. "Industrielle Souveränität").
### OUTPUT FORMAT(Strict JSON):
{
"briefingSummary": string,
"designVision": string
}
`;
const p3Resp = await axios.post('https://openrouter.ai/api/v1/chat/completions', {
model: 'google/gemini-3-flash-preview',
messages: [
{ role: 'system', content: pass3SystemPrompt },
{ role: 'user', content: `BRIEFING(TRUTH SOURCE): \n${briefing} \n\nEXISTING WEBSITE(CONTEXT): \n${distilledCrawl} \n\nEXTRACTED FACTS: \n${JSON.stringify(facts, null, 2)} ` }
],
response_format: { type: 'json_object' }
}, { headers: { 'Authorization': `Bearer ${apiKey} `, 'Content-Type': 'application/json' } });
addUsage(p3Resp.data);
if (!p3Resp.data.choices?.[0]?.message?.content) {
console.error('❌ Pass 3 failed. Response:', JSON.stringify(p3Resp.data, null, 2));
throw new Error('Pass 3: No content in response');
}
const strategy = JSON.parse(cleanJson(p3Resp.data.choices[0].message.content));
// 4. PASS 4: Information Architecture (Sitemap)
console.log(' ↳ Pass 4: Information Architecture...');
const pass4SystemPrompt = `
You are a Senior UX Architect.Design a hierarchical sitemap following the 'Industrial Logic' principle.
EVERYTHING MUST BE IN GERMAN.
### SITEMAP RULES:
1. ** HIERARCHY **: Build a logical tree.Group by category(e.g., "Kern-Präsenz", "Lösungen", "Vertrauen", "Rechtliches").
2. ** INTENT **: Each page MUST have a title and a brief functional conversion intent(desc).
3. ** COMPREHENSIVENESS **: Ensure all 'selectedPages' and 'features' from Pass 1 are represented.
4. ** LANGUAGE **: STRICT GERMAN TITLES.Do NOT use "Home", "About", "Services".Use "Startseite", "Über uns", "Leistungen".
5. ** NO IMPLEMENTATION NOTES **: Do NOT add implementation details in parentheses to titles (e.g. NO "Startseite (Hero-Video)", NO "About (Timeline)"). Keep titles clean and abstract.
### DATA CONTEXT:
${JSON.stringify({ facts, strategy }, null, 2)}
### OUTPUT FORMAT(Strict JSON):
{
"websiteTopic": string,
"sitemap": [{ "category": string, "pages": [{ "title": string, "desc": string }] }]
}
`;
const p4Resp = await axios.post('https://openrouter.ai/api/v1/chat/completions', {
model: 'google/gemini-3-flash-preview',
messages: [{ role: 'system', content: pass4SystemPrompt }, { role: 'user', content: `BRIEFING(TRUTH SOURCE): \n${briefing} \n\nDISTILLED_CRAWL(CONTEXT): \n${distilledCrawl} ` }],
response_format: { type: 'json_object' }
}, { headers: { 'Authorization': `Bearer ${apiKey} `, 'Content-Type': 'application/json' } });
addUsage(p4Resp.data);
if (!p4Resp.data.choices?.[0]?.message?.content) {
console.error('❌ Pass 4 failed. Response:', JSON.stringify(p4Resp.data, null, 2));
throw new Error('Pass 4: No content in response');
}
const ia = JSON.parse(cleanJson(p4Resp.data.choices[0].message.content));
// 5. PASS 5: Position Synthesis & Pricing Transparency
console.log(' ↳ Pass 5: Position Synthesis...');
// Determine which positions are actually relevant to avoid hallucinations
const requiredPositions = [
"1. Das technische Fundament",
facts.selectedPages.length + facts.otherPages.length > 0 ? "2. Individuelle Seiten" : null,
facts.features.length > 0 ? "3. System-Module (Features)" : null,
facts.functions.length > 0 ? "4. Logik-Funktionen" : null,
facts.apiSystems.length > 0 ? "5. Schnittstellen (API)" : null,
facts.cmsSetup ? "6. Inhaltsverwaltung (CMS)" : null,
"7. Inszenierung & Interaktion", // Always include for high-end strategy
facts.multilang ? "8. Mehrsprachigkeit" : null,
"9. Inhaltliche Initial-Pflege",
"10. Sorglos-Paket (Betrieb & Pflege)"
].filter(Boolean);
const pass5SystemPrompt = `
You are a Senior Solution Architect. Your goal is ABSOLUTE TRANSPARENCY and professionalism.
Each position in the quote must be perfectly justified and detailed using an objective, technical tone.
### REQUIRED POSITION TITLES (STRICT - ONLY DESCRIBE THESE):
${requiredPositions.map(p => `"${p}"`).join(", ")}
### MAPPING RULES (STRICT):
- ** 1. Das technische Fundament **: Infrastructure, Hosting setup, SEO-Basics, Analytics, Environments.
- ** 2. Individuelle Seiten **: Layout / structure for specific pages. ** RULE **: If quantity is high (e.g. > 10), lead with "Umsetzung von [QTY] individuellen Einzelseiten...".
- ** 3. System-Module (Features) **: Functional systems like Blog, News, Products, Jobs, References. ** RULE **: Describe exactly 1 thing if qty is 1. If qty is 0, DO NOT DESCRIBE THIS.
- ** 4. Logik-Funktionen **: Logic modules like Search, Filter, Forms, PDF-Export.
- ** 5. Schnittstellen (API) **: Data Syncs with CRM, ERP, Payment systems.
- ** 6. Inhaltsverwaltung (CMS) **: Setup and mapping for CMS.
- ** 7. Inszenierung & Interaktion **: Hero-stories, visual effects, configurators.
- ** 8. Mehrsprachigkeit **: Architecture scaling for multiple languages.
- ** 9. Inhaltliche Initial-Pflege **: Manual data entry / cleanup.
- ** 10. Sorglos-Paket (Betrieb & Pflege) **: ** RULE **: Describe as "1 Jahr Sicherung des technischen Betriebs, Instandhaltung, Sicherheits-Updates und Inhalts-Aktualisierungen gemäß AGB Punkt 7a."
### RULES FOR positionDescriptions(STRICT):
1. ** ABSOLUTE RULE: NO FIRST PERSON **: NEVER use "Ich", "Mein", "Wir" or "Unser". Lead with nouns or passive verbs.
2. ** QUANTITY PARITY (ULTRA-STRICT) **: The description MUST list EXACTLY the number of items matching the 'qty' for that position. If qty is 3, describe exactly 3 items. If qty is 1, describe exactly 1 item. Do NOT "stuff" additional features into one description.
3. ** LOGIC GUARD (CMS) **: If 'cmsSetup' is false in the DATA CONTEXT, you MUST NOT mention "CMS", "Modul", "Management System" or "Inhaltsverwaltung". Use "Statische Seite" or "Darstellung".
4. ** STATIC vs DYNAMIC **: If no complex logic was extracted in Pass 2 for a feature, describe it as a technical layout/page, not as a system.
5. ** PROFESSIONAL TONE **: Use "Erstellung von...", "Anbindung der...", "Implementierung technischer...", "Bereitstellung von...".
6. ** CONCISE & ITEM-BASED **: Use technical, high-density sentences. Name specific industry terms from context.
7. ** ITEMIZED SYNTHESIS **: Mention EVERY component selected in Pass 1.
8. ** HARD SPECIFICS **: If the briefing mentions "Glasfaser-Trassen" or "Schwerlast-Logistik", IT MUST BE IN THE DESCRIPTION.
9. ** INDUSTRIAL AMBITION **: Describe it as a high-end technical solution. Avoid "schöne Website" or marketing fluff.
10. ** PAGES **: For "2. Individuelle Seiten", list the pages. ** ABSOLUTE RULE **: Do NOT add implementation details or technical notes in parentheses (e.g. NO "(Matrix-Struktur)", NO "(Timeline-Modul)"). Use clean titles like "Startseite, Über uns, Leistungen".
11. ** LOGIC **: Describe the ACTUAL logic (e.g., "Volltextsuche mit Auto-Complete", not "eine Suche").
12. ** KEYS **: Return EXACTLY the keys defined in "POSITION TITLES".
13. ** NO AGB **: NEVER mention "AGB" or "Geschäftsbedingungen".
### EXAMPLES(PASSIVE & TECHNICAL):
- ** GOOD **: "Konfiguration der CMS-Infrastruktur zur unabhängigen Verwaltung von Produkt-Katalogen und News-Beiträgen."
- ** GOOD **: "Implementierung einer Volltextsuche inkl. Kategorisierungs-Logik für effizientes Auffinden von Projektreferenzen."
- ** GOOD **: "Native API-Anbindung an das ERP-System zur Echtzeit-Synchronisation von Bestandsdaten."
- ** BAD **: "Ich richte dir das CMS ein."
- ** BAD **: "Ich programmiere eine tolle Suche für deine Seite."
### DATA CONTEXT:
${JSON.stringify({ facts, details, strategy, ia }, null, 2)}
### OUTPUT FORMAT(Strict JSON):
{
"positionDescriptions": { "1. Das technische Fundament": string, ... }
}
`;
const p5Resp = await axios.post('https://openrouter.ai/api/v1/chat/completions', {
model: 'google/gemini-3-flash-preview',
messages: [{ role: 'system', content: pass5SystemPrompt }, { role: 'user', content: briefing }],
response_format: { type: 'json_object' }
}, { headers: { 'Authorization': `Bearer ${apiKey} `, 'Content-Type': 'application/json' } });
addUsage(p5Resp.data);
if (!p5Resp.data.choices?.[0]?.message?.content) {
console.error('❌ Pass 5 failed. Response:', JSON.stringify(p5Resp.data, null, 2));
throw new Error('Pass 5: No content in response');
}
const positionsData = JSON.parse(cleanJson(p5Resp.data.choices[0].message.content));
// 6. PASS 6: The Industrial Critic
console.log(' ↳ Pass 6: The Industrial Critic (Quality Gate)...');
const pass6SystemPrompt = `
You are the "Industrial Critic".Your goal is to catch quality regressions and ensure the document is bespoke, technical, and professional.
Analyze the CURRENT_STATE against the BRIEFING_TRUTH.
### CRITICAL ERROR CHECKLIST(FAIL IF FOUND):
1. ** Hallucination Leakage **: FAIL if names of people(e.g., "Frieder Helmich"), specific software versions, or invented details are used unless they appear EXACTLY in the BRIEFING.
- ** CRITICAL **: Forbid "Sie", "Ansprechpartner" or "Unternehmen" for personName if a name IS in the briefing.If none is in briefing, use empty string.
2. ** Logic Conflict **: FAIL if isRelaunch is true but briefingSummary claims no website exists.
- FAIL if the description in positionDescriptions mentions more items than extracted in facts.
3. ** Implementation Fluff **: FAIL if tech - stack details are mentioned(React, etc.).Focus on Concept & Result.
4. ** Genericism Check(CRITICAL) **: FAIL if any text sounds like it could apply to ANY company.It MUST mention specific industry details(e.g., "Kabeltiefbau", "Infrastruktur-Zentrum") from the Briefing or Crawl.
6. ** Namen-Verbot (STRICT) **: FAIL if any personal names (e.g. "Danny Joseph", "Joseph", etc.) appear in 'briefingSummary' or 'designVision'. Use abstract terms like "Unternehmensführung" or "Management" if necessary.
7. ** LOGIC GUARD (CMS) **: If 'cmsSetup' is false in the DATA CONTEXT, FAIL if any 'positionDescriptions' or 'briefingSummary' mentions "CMS", "Content Management System" or "Inhaltsverwaltung".
8. ** AGB BAN **: FAIL if "Allgemeine Geschäftsbedingungen" or "AGB" appear anywhere.
9. ** Length Check **: Briefing (ca. 6 Sätze) und Vision (ca. 4 Sätze). Kürze Texte, die zu ausschweifend sind, auf das Wesentliche.
### MISSION:
Return updated fields ONLY.Specifically focus on hardening 'positionDescriptions', 'sitemap', 'briefingSummary', and 'designVision'.
### DATA CONTEXT:
${JSON.stringify({ facts, strategy, ia, positionsData }, null, 2)}
`;
const p6Resp = await axios.post('https://openrouter.ai/api/v1/chat/completions', {
model: 'google/gemini-3-flash-preview',
messages: [{ role: 'system', content: pass6SystemPrompt }, { role: 'user', content: `BRIEFING_TRUTH: \n${briefing} ` }],
response_format: { type: 'json_object' }
}, { headers: { 'Authorization': `Bearer ${apiKey} `, 'Content-Type': 'application/json' } });
addUsage(p6Resp.data);
if (!p6Resp.data.choices?.[0]?.message?.content) {
console.error('❌ Pass 6 failed. Response:', JSON.stringify(p6Resp.data, null, 2));
throw new Error('Pass 6: No content in response');
}
const reflection = JSON.parse(cleanJson(p6Resp.data.choices[0].message.content));
// 6. Reflection Merge Utility
const mergeReflection = (state: any, reflection: any) => {
let result = { ...state };
const unwrap = (obj: any): any => {
if (!obj || typeof obj !== 'object' || Array.isArray(obj)) return obj;
// Always unwrap "0" if it exists, regardless of other keys (AI often nests)
if (obj["0"]) return unwrap(obj["0"]);
if (obj.state && Object.keys(obj).length === 1) return unwrap(obj.state);
if (obj.facts && Object.keys(obj).length === 1) return unwrap(obj.facts);
return obj;
};
const cleanedReflection = unwrap(reflection);
Object.entries(cleanedReflection).forEach(([key, value]) => {
if (value && value !== "" && value !== "null") {
result[key] = value;
}
});
return result;
};
let finalState = mergeReflection({
...initialState,
...facts,
...strategy,
...ia,
...positionsData
}, reflection);
finalState.statusQuo = facts.isRelaunch ? 'Relaunch' : 'Neuentwicklung';
// Recipient Mapping
if (finalState.personName) finalState.name = finalState.personName;
if (finalState.email) finalState.email = finalState.email;
// Normalization Layer: Map hallucinated German keys back to internal keys
const normalizationMap: Record<string, string> = {
"Briefing-Zusammenfassung": "briefingSummary",
"Design-Vision": "designVision",
"Zusammenfassung": "briefingSummary",
"Vision": "designVision",
"BRIEFING_SUMMARY": "briefingSummary",
"DESIGN_VISION": "designVision"
};
Object.entries(normalizationMap).forEach(([gerKey, intKey]) => {
if (finalState[gerKey] && !finalState[intKey]) {
if (typeof finalState[gerKey] === 'object' && !Array.isArray(finalState[gerKey])) {
finalState[intKey] = Object.values(finalState[gerKey]).join('\n\n');
} else {
finalState[intKey] = finalState[gerKey];
}
}
});
// Final Logic Guard: Strictly strip CMS from ALL descriptions and fields if not enabled
if (!finalState.cmsSetup) {
const stripCMS = (obj: any): any => {
if (typeof obj === 'string') {
return obj.replace(/CMS|Content-Management-System|Inhaltsverwaltung/gi, 'Plattform-Struktur');
}
if (Array.isArray(obj)) {
return obj.map(stripCMS);
}
if (obj !== null && typeof obj === 'object') {
const newObj: any = {};
Object.entries(obj).forEach(([k, v]) => {
newObj[k] = stripCMS(v);
});
return newObj;
}
return obj;
};
finalState = stripCMS(finalState);
}
// Sitemap Normalization (German keys to internal)
if (Array.isArray(finalState.sitemap)) {
finalState.sitemap = finalState.sitemap.map((cat: any) => ({
category: cat.category || cat.kategorie || cat.Kategorie || cat.title || "Allgemein",
pages: (cat.pages || cat.seiten || cat.Seiten || []).map((page: any) => ({
title: page.title || page.titel || page.Titel || "Seite",
desc: page.desc || page.beschreibung || page.Beschreibung || page.description || ""
}))
}));
}
// Position Descriptions Normalization (Strict Title Mapping + Index-based Fallback)
if (finalState.positionDescriptions) {
const normalized: Record<string, string> = {};
const rawPositions = finalState.positionDescriptions;
// 1. Initial cleanup
Object.entries(rawPositions).forEach(([key, value]) => {
const normalizedValue = typeof value === 'object' ? (value as any).beschreibung || (value as any).description || JSON.stringify(value) : value;
normalized[key] = normalizedValue as string;
});
// 2. Index-based matching (Map "10. Foo" to "10. Bar")
const standardTitles = [
"1. Das technische Fundament",
"2. Individuelle Seiten",
"3. System-Module (Features)",
"4. Logik-Funktionen",
"5. Schnittstellen (API)",
"6. Inhaltsverwaltung (CMS)",
"7. Inszenierung & Interaktion",
"8. Mehrsprachigkeit",
"9. Inhaltliche Initial-Pflege",
"10. Sorglos-Paket (Betrieb & Pflege)"
];
standardTitles.forEach(std => {
const prefix = std.split('.')[0] + '.'; // e.g., "10."
// Find any key in the AI output that starts with this number
const matchingKey = Object.keys(normalized).find(k => k.trim().startsWith(prefix));
if (matchingKey && matchingKey !== std) {
normalized[std] = normalized[matchingKey];
// Keep the old key too just in case, but prioritize the standard one
}
});
finalState.positionDescriptions = normalized;
}
// Normalize final state
if (Array.isArray(finalState.positionDescriptions)) {
const normalized: Record<string, string> = {};
finalState.positionDescriptions.forEach((item: any) => {
const key = item.feature || item.id || item.title || item.pos;
if (key) normalized[key] = item.description || item.desc;
});
finalState.positionDescriptions = normalized;
}
if (finalState.sitemap && !Array.isArray(finalState.sitemap)) {
if (finalState.sitemap.categories) finalState.sitemap = finalState.sitemap.categories;
else if (finalState.sitemap.sitemap) finalState.sitemap = finalState.sitemap.sitemap;
else {
const entries = Object.entries(finalState.sitemap);
if (entries.every(([_, v]) => Array.isArray(v))) {
finalState.sitemap = entries.map(([category, pages]) => ({ category, pages }));
}
}
}
// Final Post-Reflection Budget Sync (Hard Pruning if still over)
if (budget) {
const targetValue = parseInt(budget.replace(/[^0-9]/g, ''));
if (!isNaN(targetValue)) {
console.log(`⚖️ Final Budget Audit(${targetValue} € target)...`);
let currentTotals = calculateTotals(finalState, PRICING);
// Step-by-step pruning if too expensive
if (currentTotals.totalPrice > targetValue) {
console.log(`⚠️ Budget exceeded(${currentTotals.totalPrice} €).Pruning scope to fit ${targetValue} €...`);
// 1. Remove optional "other" stuff
finalState.otherFeatures = [];
finalState.otherFunctions = [];
finalState.otherTech = [];
// 2. Remove non-critical functions
const funcPriority = ['search', 'filter', 'calendar', 'multilang'];
for (const f of funcPriority) {
if (currentTotals.totalPrice <= targetValue) break;
if (finalState.functions.includes(f)) {
finalState.functions = finalState.functions.filter((x: string) => x !== f);
currentTotals = calculateTotals(finalState, PRICING);
}
}
// 3. Remove least critical features if still over
const featurePriority = ['events', 'blog_news', 'products'];
for (const p of featurePriority) {
if (currentTotals.totalPrice <= targetValue) break;
if (finalState.features.includes(p)) {
finalState.features = finalState.features.filter((f: string) => f !== p);
currentTotals = calculateTotals(finalState, PRICING);
}
}
// 4. Reduce page count (Selected Pages AND Sitemap)
while (currentTotals.totalPrice > targetValue && (finalState.selectedPages.length > 4 || currentTotals.totalPagesCount > 5)) {
if (finalState.selectedPages.length > 4) {
finalState.selectedPages.pop();
}
// Prune Sitemap to match
if (finalState.sitemap && Array.isArray(finalState.sitemap)) {
const lastCat = finalState.sitemap[finalState.sitemap.length - 1];
if (lastCat && lastCat.pages && lastCat.pages.length > 0) {
lastCat.pages.pop();
if (lastCat.pages.length === 0) finalState.sitemap.pop();
}
}
currentTotals = calculateTotals(finalState, PRICING);
}
// 5. Final fallback: Remove second feature if still over
if (currentTotals.totalPrice > targetValue && finalState.features.length > 1) {
finalState.features.pop();
currentTotals = calculateTotals(finalState, PRICING);
}
}
console.log(`✅ Final budget audit complete: ${currentTotals.totalPrice}`);
}
}
return { state: finalState, usage };
}
main().catch(console.error);