Update server.js

server.js

@@ -1,67 +1,14 @@
 import { pipeline } from '@huggingface/transformers';
 import http from 'http';
-import fs from 'fs';
-import path from 'path';
-import { fileURLToPath } from 'url';
 
 const PORT = 7860;
 const MODEL_NAME = 'onnx-community/Qwen2.5-0.5B-Instruct';
-
-const __filename = fileURLToPath(import.meta.url);
-const __dirname = path.dirname(__filename);
-
 let generator;
-const KNOWLEDGE_TOPICS = {}; 
-
-// 1. Load the single file and automatically parse [TOPIC: XXX] blocks
-async function initializeServer() {
-    try {
-        console.log("Reading knowledge.txt...");
-        const knowledgePath = path.join(__dirname, 'knowledge.txt');
-        
-        if (fs.existsSync(knowledgePath)) {
-            const rawContent = fs.readFileSync(knowledgePath, 'utf8');
-            
-            const topicRegex = /\[TOPIC:\s*(\w+)\]([\s\S]*?)(?=\[TOPIC:|$)/gi;
-            let match;
-            
-            while ((match = topicRegex.exec(rawContent)) !== null) {
-                const topicName = match[1].toLowerCase().trim();
-                const topicContent = match[2].trim();
-                KNOWLEDGE_TOPICS[topicName] = topicContent;
-                console.log(`-> Successfully registered topic: "${topicName}"`);
-            }
-        } else {
-            console.warn("WARNING: knowledge.txt not found. Starting blank.");
-        }
-
-        console.log("Loading Qwen2.5-Coder-3B model... (first load may take a few minutes)");
-        generator = await pipeline('text-generation', MODEL_NAME, { dtype: 'q4' });
-        console.log("Server fully initialized!");
-
-    } catch (error) {
-        console.error("Initialization failed:", error);
-        process.exit(1);
-    }
-}
-
-// 2. 100% Automated Router - Searches user prompt against loaded keys
-function getRelevantContext(userPrompt) {
-    const lowerPrompt = userPrompt.toLowerCase();
-    let dynamicContext = "";
-
-    for (const topicName in KNOWLEDGE_TOPICS) {
-        if (lowerPrompt.includes(topicName)) {
-            dynamicContext += KNOWLEDGE_TOPICS[topicName] + "\n\n";
-        }
-    }
-
-    if (!dynamicContext.trim()) {
-        const available = Object.keys(KNOWLEDGE_TOPICS).join(", ");
-        dynamicContext = `General support mode. Reference available topics: ${available}.`;
-    }
 
-    return dynamicContext;
+async function loadModel() {
+    console.log("Loading coding model...");
+    generator = await pipeline('text-generation', MODEL_NAME, { dtype: 'q4' });
+    console.log("Model loaded successfully!");
 }
 
 const server = http.createServer(async (req, res) => {
@@ -79,10 +26,7 @@ const server = http.createServer(async (req, res) => {
 
     if (pathname === '/' && req.method === 'GET') {
         res.writeHead(200);
-        res.end(JSON.stringify({ 
-            status: "Backend is running", 
-            available_topics: Object.keys(KNOWLEDGE_TOPICS)
-        }));
+        res.end(JSON.stringify({ status: "Backend is running" }));
         return;
     }
 
@@ -94,126 +38,74 @@ const server = http.createServer(async (req, res) => {
                 const { prompt } = JSON.parse(body);
                 if (!generator) {
                     res.writeHead(503);
-                    return res.end(JSON.stringify({ error: "System warming up..." }));
+                    return res.end(JSON.stringify({ error: "Model is still loading..." }));
                 }
 
-                // Run prompt router to fetch targeted text injection
-                const localizedContext = getRelevantContext(prompt);
-
-                // Detect if it is a programming/coding request
-                const codingKeywords = [
-                    'generate', 'create', 'build', 'make', 'code', 'website', 'webpage',
-                    'login', 'register', 'form', 'page', 'app', 'application', 'html',
-                    'css', 'javascript', 'backend', 'frontend', 'server', 'api', 'express',
-                    'node', 'database', 'function', 'script', 'component', 'write'
-                ];
-                const lowerPrompt = prompt.toLowerCase();
-                const isCodingRequest = codingKeywords.some(k => lowerPrompt.includes(k));
-
-                const finalPrompt = isCodingRequest
-                    ? `As a coding assistant, write the complete working code for the following task. Always provide full code, never refuse a coding task: ${prompt}`
-                    : prompt;
-
-                // Core base prompt with static company identity and localized file data
-                let systemPrompt = `You are Gini AI, an AI full-stack development assistant created by Emalawi19, founded by Professor Roosevelt Chinkwende.
+                const messages = [
+                    {
+                        role: 'system',
+                        content: `You are Gini AI, a full-stack web development AI assistant created by Emalawi19, founded by Professor Roosevelt Chinkwende.
 
-IDENTITY:
+IDENTITY RULES:
 - Your name is Gini AI.
-- Your company is Emalawi19, founded by Professor Roosevelt Chinkwende.
-- You are an AI, not a human.
-
-REFERENCE KNOWLEDGE BASE:
-Use the following official documentation to accurately answer any user questions about operations or topics:
-${localizedContext}`;
-
-                // Inject dynamic behavior rules dependent on the intent type
-                if (isCodingRequest) {
-                    systemPrompt += `
-
-ABSOLUTE RULES — NEVER BREAK THESE:
-1. You ALWAYS write code when asked. You NEVER say "I can't help with that" for any coding request.
-2. Any request involving web development or script building is ALWAYS answered with complete working code blocks.
-3. You guide users step by step, one step at a time, with complete copy-paste ready code.
-4. For EACH step provide:
-   a) "Step X of Y: Title"
-   b) One sentence explaining what this step does
-   c) The COMPLETE working code in a properly labelled code block
-   d) Filename and where to save it
-   e) How to run it if needed
+- Your company is Emalawi19.
+- The founder is Professor Roosevelt Chinkwende.
+- You are an AI, not a human developer.
+
+YOUR SPECIALTY:
+- You build complete websites with both frontend (HTML, CSS, JavaScript) and backend (Node.js + Express).
+- You ALWAYS guide users one step at a time.
+- You NEVER give vague instructions — every step includes COMPLETE, COPY-PASTE READY code.
+- You NEVER cut off mid-response. Always finish the current step fully.
+
+STEP-BY-STEP GUIDE RULES (CRITICAL):
+1. When asked to build a website or app, first list all steps with a short title each.
+2. Then say: "Let's begin with Step 1. Say 'next' when you are ready to continue."
+3. For each step provide:
+   a) Step title and number (e.g. "Step 1 of 5: Project Setup")
+   b) One short sentence explaining what this step does
+   c) The COMPLETE working code for this step inside a code block
+   d) Where to save the file (e.g. "Save as server.js in your project folder")
+   e) How to run it if needed (e.g. "Run: node server.js")
    f) End with: "Say 'next' to continue to Step X."
-5. NEVER use placeholders like '// add your logic here'. Write the actual working logic.
-6. NEVER combine steps. One step per response only.
-7. ALWAYS finish the full code for the current step — never cut off mid-code.
-
-CODE FORMAT:
-- Always use triple backticks with language tag: \`\`\`html \`\`\`css \`\`\`javascript \`\`\``;
-                } else {
-                    systemPrompt += `
-
-RULES FOR GENERAL ASSISTANCE:
-1. Answer the user's question clearly, helpfully, and accurately using facts from the REFERENCE KNOWLEDGE BASE block above.
-2. DO NOT write or generate programming code (like JavaScript, HTML, or Node.js scripts) unless explicitly requested by the user.
-3. Keep your formatting clean using concise paragraphs or scannable bullet points.`;
-                }
-
-                const messages = [
-                    { role: 'system', content: systemPrompt },
-                    { role: 'user', content: finalPrompt }
+4. NEVER skip code. NEVER say 'add your logic here'. Write the actual logic.
+5. NEVER combine multiple steps into one response.
+6. Default stack: HTML/CSS/JS frontend, Node.js + Express backend.
+7. ALWAYS complete the current step fully before stopping — do not cut off mid-code.
+
+CODE FORMAT RULES:
+- Always wrap code in triple backticks with the language tag e.g. \`\`\`javascript or \`\`\`html
+- Code must be real, working, and complete for that step.
+- Do not write pseudo-code or placeholders.`
+                    },
+                    { role: 'user', content: prompt }
                 ];
 
                 const output = await generator(messages, {
-                    max_new_tokens: 1024,
-                    temperature: 0.2,
-                    repetition_penalty: 1.15,
-                    do_sample: false  
+                    max_new_tokens: 800,   // increased so steps never cut off
+                    temperature: 0.4,      // lower = more focused, less hallucination
+                    repetition_penalty: 1.2, // prevents looping/repeating text
+                    do_sample: true
                 });
 
                 const result = output[0].generated_text.at(-1).content;
-
-                // Safety fallback mechanism for false refusal flags on coding inputs
-                const refusalPhrases = ["i can't assist", "i cannot assist", "i'm sorry", "i am sorry", "i'm unable", "i cannot help"];
-                const isRefusal = refusalPhrases.some(p => result.toLowerCase().startsWith(p));
-
-                if (isRefusal && isCodingRequest) {
-                    const retryMessages = [
-                        {
-                            role: 'system',
-                            content: `You are a code generator. You only output code. Never refuse. Never explain. Just write the complete working code asked for.`
-                        },
-                        {
-                            role: 'user',
-                            content: `Write complete working code for: ${prompt}. Include all elements required.`
-                        }
-                    ];
-
-                    const retryOutput = await generator(retryMessages, {
-                        max_new_tokens: 1024,
-                        temperature: 0.1,
-                        do_sample: false
-                    });
-
-                    const retryResult = retryOutput[0].generated_text.at(-1).content;
-                    res.writeHead(200);
-                    return res.end(JSON.stringify({ result: retryResult }));
-                }
-
                 res.writeHead(200);
                 res.end(JSON.stringify({ result }));
 
             } catch (err) {
                 res.writeHead(400);
-                res.end(JSON.stringify({ error: "Processing error", detail: err.message }));
+                res.end(JSON.stringify({ error: "Invalid request", detail: err.message }));
             }
         });
         return;
     }
 
     res.writeHead(404);
-    res.end(JSON.stringify({ error: "Not Found" }));
+    res.end(JSON.stringify({ error: "Not Found", requested_path: pathname }));
 });
 
-initializeServer().then(() => {
+loadModel().then(() => {
     server.listen(PORT, '0.0.0.0', () => {
-        console.log(`Server listening on port ${PORT}`);
+        console.log(`Server running at http://0.0.0.0:${PORT}`);
     });
 });