Update server.js

server.js

@@ -3,20 +3,20 @@ import http from 'http';
 import fs from 'fs';
 import path from 'path';
 
-const PORT       = 7860;
-const MODEL_NAME = 'onnx-community/Qwen2.5-0.5B-Instruct';
+const PORT          = 7860;
+const MODEL_NAME    = 'onnx-community/Qwen2.5-0.5B-Instruct';
 const KNOWLEDGE_DIR = './knowledge';
 let generator;
-let knowledgeBase = [];
+let knowledgeBase   = [];
 
-// ── KNOWLEDGE FILES ───────────────────────────────────────────────────────────
+// ── KNOWLEDGE ─────────────────────────────────────────────────────────────────
 function loadKnowledge() {
     if (!fs.existsSync(KNOWLEDGE_DIR)) { fs.mkdirSync(KNOWLEDGE_DIR); return; }
     const files = fs.readdirSync(KNOWLEDGE_DIR).filter(f => f.endsWith('.txt'));
     knowledgeBase = [];
     for (const file of files) {
         const content = fs.readFileSync(path.join(KNOWLEDGE_DIR, file), 'utf-8');
-        const chunks  = splitChunks(content, 600, 60);
+        const chunks  = splitChunks(content, 800, 80);
         chunks.forEach(c => knowledgeBase.push({ source: file, text: c }));
     }
     console.log(`Knowledge loaded: ${files.length} files, ${knowledgeBase.length} chunks`);
@@ -24,7 +24,7 @@ function loadKnowledge() {
 
 function splitChunks(text, size, overlap) {
     const chunks = [];
-    let start = 0;
+    let start    = 0;
     while (start < text.length) {
         chunks.push(text.slice(start, start + size));
         start += size - overlap;
@@ -32,19 +32,32 @@ function splitChunks(text, size, overlap) {
     return chunks;
 }
 
-function retrieveContext(prompt, topK = 4) {
+function retrieveContext(prompt, topK = 5) {
     if (knowledgeBase.length === 0) return '';
-    const words  = prompt.toLowerCase().split(/\W+/).filter(w => w.length > 2);
-    const scored = knowledgeBase.map(chunk => ({
-        ...chunk,
-        score: words.reduce((acc, w) => acc + (chunk.text.toLowerCase().includes(w) ? 1 : 0), 0)
-    }));
-    return scored
+
+    const words = prompt.toLowerCase()
+        .split(/\W+/)
+        .filter(w => w.length > 2);
+
+    const scored = knowledgeBase.map(chunk => {
+        const lower = chunk.text.toLowerCase();
+        let score   = 0;
+        for (const w of words) {
+            // Count every occurrence not just presence — better scoring
+            const matches = (lower.match(new RegExp(w, 'g')) || []).length;
+            score += matches;
+        }
+        return { ...chunk, score };
+    });
+
+    const top = scored
         .filter(c => c.score > 0)
         .sort((a, b) => b.score - a.score)
-        .slice(0, topK)
-        .map(c => `[${c.source}]\n${c.text}`)
-        .join('\n\n---\n\n');
+        .slice(0, topK);
+
+    if (top.length === 0) return '';
+
+    return top.map(c => c.text).join('\n\n---\n\n');
 }
 
 // ── MODEL ─────────────────────────────────────────────────────────────────────
@@ -56,22 +69,19 @@ async function loadModel() {
 
 async function generateResponse(messages) {
     const output = await generator(messages, {
-        max_new_tokens:     500,
-        temperature:        0.3,
-        repetition_penalty: 1.2,
+        max_new_tokens:     600,
+        temperature:        0.2,
+        repetition_penalty: 1.15,
         do_sample:          false
     });
-    // Extract only the assistant reply content
     const generated = output[0].generated_text;
-    if (Array.isArray(generated)) {
-        return generated.at(-1)?.content || '';
-    }
+    if (Array.isArray(generated)) return generated.at(-1)?.content || '';
     return String(generated || '');
 }
 
 // ── SERVER ────────────────────────────────────────────────────────────────────
 const server = http.createServer(async (req, res) => {
-    res.setHeader('Access-Control-Allow-Origin', '*');
+    res.setHeader('Access-Control-Allow-Origin',  '*');
     res.setHeader('Access-Control-Allow-Methods', 'GET, POST, OPTIONS');
     res.setHeader('Access-Control-Allow-Headers', 'Content-Type');
 
@@ -79,18 +89,16 @@ const server = http.createServer(async (req, res) => {
 
     const pathname = req.url.split('?')[0];
 
-    // Status check
     if (pathname === '/' && req.method === 'GET') {
         res.setHeader('Content-Type', 'application/json');
         res.writeHead(200);
         return res.end(JSON.stringify({
             status: "running",
-            model: MODEL_NAME,
+            model:  MODEL_NAME,
             knowledge_chunks: knowledgeBase.length
         }));
     }
 
-    // Reload knowledge
     if (pathname === '/reload-knowledge' && req.method === 'POST') {
         loadKnowledge();
         res.setHeader('Content-Type', 'application/json');
@@ -98,57 +106,72 @@ const server = http.createServer(async (req, res) => {
         return res.end(JSON.stringify({ message: `Reloaded: ${knowledgeBase.length} chunks` }));
     }
 
-    // Main generate endpoint — returns plain JSON (no SSE, no streaming delays)
     if (pathname === '/generate' && req.method === 'POST') {
         let body = '';
         req.on('data', c => { body += c.toString(); });
         req.on('end', async () => {
             res.setHeader('Content-Type', 'application/json');
-
             try {
-                const { prompt, system } = JSON.parse(body);
+                const { prompt } = JSON.parse(body);
 
                 if (!generator) {
                     res.writeHead(503);
-                    return res.end(JSON.stringify({ error: "Model still loading, please wait..." }));
+                    return res.end(JSON.stringify({ error: "Model still loading..." }));
                 }
 
-                // RAG: retrieve relevant knowledge chunks
-                const ragContext = retrieveContext(prompt, 4);
-                const ragSection = ragContext
-                    ? `\n\nKNOWLEDGE BASE — use ONLY this information to answer:\n${ragContext}\n`
-                    : '';
+                console.log(`Query: "${prompt}"`);
 
-                // Build final system prompt
-                const finalSystem = (system || `You are Mlimi Connect AI, a free agricultural advisor for Malawian farmers. Only answer agriculture questions.`) + ragSection;
+                // ── RAG: find relevant knowledge chunks ───────────────────
+                const ragContext = retrieveContext(prompt, 5);
+                console.log(`RAG chunks found: ${ragContext.length} chars`);
+
+                // ── Build system prompt with knowledge injected ────────────
+                // IMPORTANT: knowledge comes FIRST, before any other instruction
+                const systemPrompt = ragContext
+                    ? `You are Mlimi Connect AI, a free agricultural advisor for Malawian farmers.
+
+KNOWLEDGE BASE — THIS IS YOUR ONLY SOURCE OF INFORMATION. USE ONLY THIS:
+===START OF KNOWLEDGE===
+${ragContext}
+===END OF KNOWLEDGE===
+
+STRICT RULES:
+1. Answer ONLY using the knowledge provided above between ===START=== and ===END===.
+2. Do NOT add information from outside the knowledge base.
+3. Do NOT be vague. Give specific details: variety names, exact spacing, fertilizer amounts, timing.
+4. Structure your answer clearly with numbered steps.
+5. If the knowledge above does not contain the answer, say: "I don't have specific information on that in my knowledge base."
+6. ONLY answer agriculture questions. For anything else say: "I can only help with farming questions."`
+
+                    : `You are Mlimi Connect AI, a free agricultural advisor for Malawian farmers.
+
+I don't have specific notes on that topic in my knowledge base yet.
+Give a brief, honest answer based on general Malawian agricultural knowledge.
+Keep it practical and specific to Malawi's conditions.
+ONLY answer agriculture questions.`;
 
                 const messages = [
-                    { role: 'system', content: finalSystem },
-                    { role: 'user',   content: prompt       }
+                    { role: 'system', content: systemPrompt },
+                    { role: 'user',   content: prompt        }
                 ];
 
-                console.log(`Generating response for: "${prompt.slice(0, 60)}..."`);
-
                 let result = await generateResponse(messages);
 
-                // Fallback if model returns empty
-                if (!result || result.trim().length < 5) {
-                    if (ragContext) {
-                        result = `Here is what I know about this topic:\n\n${ragContext.slice(0, 600)}`;
-                    } else {
-                        result = "I don't have specific information on that topic. Please consult your local agricultural extension officer (AEO) for advice.";
-                    }
+                // If still empty, return the raw knowledge chunk directly
+                if (!result || result.trim().length < 10) {
+                    result = ragContext
+                        ? `Here is what my knowledge base says:\n\n${ragContext.slice(0, 800)}`
+                        : "I don't have specific information on that topic. Please ask your local agricultural extension officer.";
                 }
 
-                console.log(`Response ready: ${result.length} chars`);
-
+                console.log(`Response: ${result.slice(0, 80)}...`);
                 res.writeHead(200);
                 res.end(JSON.stringify({ result }));
 
             } catch (err) {
-                console.error("Generation error:", err.message);
+                console.error("Error:", err.message);
                 res.writeHead(500);
-                res.end(JSON.stringify({ error: err.message || "Generation failed" }));
+                res.end(JSON.stringify({ error: err.message }));
             }
         });
         return;
@@ -162,6 +185,6 @@ const server = http.createServer(async (req, res) => {
 loadKnowledge();
 loadModel().then(() => {
     server.listen(PORT, '0.0.0.0', () => {
-        console.log(`Mlimi Connect backend running on port ${PORT}`);
+        console.log(`Mlimi Connect backend on port ${PORT}`);
     });
 });