Rework algorithm

lib/llm.ts

@@ -3,6 +3,7 @@ import { fileURLToPath } from "node:url";
 
 import {
   type ChatHistoryItem,
+  type LLamaChatPromptOptions,
   LlamaChatSession,
   type LlamaModel,
   createModelDownloader,
@@ -19,7 +20,7 @@ export async function getModel(model: string) {
   });
   const modelPath = await downloader.download();
   const llama = await getLlama({
-    maxThreads: 2,
+    maxThreads: 6,
   });
   return await llama.loadModel({ modelPath });
 }
@@ -29,7 +30,11 @@ export type Message = {
   text: string;
 };
 
-export async function complete(model: LlamaModel, messages: Message[]) {
+export async function complete(
+  model: LlamaModel,
+  messages: Message[],
+  options: LLamaChatPromptOptions = {},
+) {
   if (messages.length < 1) throw new Error("messages are empty");
   const init = messages.slice(0, -1);
   const last = messages.at(-1) as Message;
@@ -61,14 +66,11 @@ export async function complete(model: LlamaModel, messages: Message[]) {
   );
 
   const res = await session.prompt(last.text, {
-    temperature: 1.0,
-    repeatPenalty: {
-      frequencyPenalty: 1,
-    },
+    trimWhitespaceSuffix: true,
     onResponseChunk(chunk) {
       process.stderr.write(chunk.text);
     },
-    maxTokens: 200,
+    ...options,
   });
   session.dispose();
   await context.dispose();