empty string fix

packages/inference-service/src/providers/llamacpp.js

@@ -66,49 +66,50 @@ async function* completeStream(prompt, options = {}) {
   if (!res.ok)
     throw new Error(`llama.cpp error: ${res.status} ${res.statusText}`);
 
-  //SSE lines can be split across network chunks, so we need to buffer them until we have a complete line
+  // SSE lines can be split across network chunks, so we buffer the incomplete
+  // tail and only process whole lines.
   let buffer = '';
 
-  function* processLine(line){
+  // Handles ONE line. Yields at most one content chunk. Never touches `lines`
+  // or the network buffer — the reader loop below owns that.
+  function* processLine(line) {
     if (!line.startsWith('data: ') || line === 'data: [DONE]') return;
 
     let json;
-
     try {
       json = JSON.parse(line.slice(6));
     } catch (err) {
-      logger.error('[llamacpp] Skipping unparseable SSE line:', line.slice(0,120));
+      logger.warn('[llamacpp] Skipping unparseable SSE line:', line.slice(0, 120));
       return;
     }
 
+    const delta = json.choices?.[0]?.delta?.content;
 
-
+    if (json.choices?.[0]?.finish_reason === 'stop') {
-    const delta = json.choices?.[0]?.delta?.content ?? '';
-
-    if (json.choices?.[0]?.finish_reason === 'stop' ) {
       finalModel = json.model ?? finalModel;
-
     }
 
     // usage arrives in a separate final chunk with empty choices array
-    if (json.usage){
+    if (json.usage) {
       finalTokenCount = (json.usage.completion_tokens ?? 0) + (json.usage.prompt_tokens ?? 0);
     }
 
-    for (const line of lines) {
+    if (delta) yield { response: delta, done: false };
-      yield* processLine(line.trim());
-    }
   }
 
   for await (const chunk of res.body) {
     buffer += Buffer.from(chunk).toString('utf-8');
 
     const lines = buffer.split('\n');
-    buffer = lines.pop() ?? ""; // keep the last line in the buffer, it may be incomplete
+    buffer = lines.pop() ?? '';  // last element may be incomplete — keep for next chunk
+
+    for (const line of lines) {
+      yield* processLine(line.trim());
+    }
   }
 
-  //Flush anything left in buffer after stream closes
+  // Flush anything left in the buffer after the stream closes
-  if (buffer.trim()){
+  if (buffer.trim()) {
     yield* processLine(buffer.trim());
   }