From 2a38ae4a955bdfd1eaa7be9571f42711b8866ed3 Mon Sep 17 00:00:00 2001
From: Wroclaw <wroclaw223@outlook.com>
Date: Sun, 19 Mar 2023 04:15:08 +0100
Subject: [PATCH] Limit sent chat to 2048 tokens.

This also solves the issue where we would request more tokens,
than the model is capable of (over 4096)
---
 package-lock.json       |  6 ++++++
 package.json            |  1 +
 src/toOpenAIMessages.ts | 30 ++++++++++++++++++------------
 src/tokenCounter.ts     |  5 +++++
 4 files changed, 30 insertions(+), 12 deletions(-)
 create mode 100644 src/tokenCounter.ts

diff --git a/package-lock.json b/package-lock.json
index 55aeb90..f5a0775 100644
--- a/package-lock.json
+++ b/package-lock.json
@@ -11,6 +11,7 @@
       "dependencies": {
         "discord.js": "^14.8.0",
         "fold-to-ascii": "^5.0.1",
+        "gpt-3-encoder": "^1.1.4",
         "openai": "^3.2.1"
       },
       "devDependencies": {
@@ -1161,6 +1162,11 @@
         "url": "https://github.com/sponsors/sindresorhus"
       }
     },
+    "node_modules/gpt-3-encoder": {
+      "version": "1.1.4",
+      "resolved": "https://registry.npmjs.org/gpt-3-encoder/-/gpt-3-encoder-1.1.4.tgz",
+      "integrity": "sha512-fSQRePV+HUAhCn7+7HL7lNIXNm6eaFWFbNLOOGtmSJ0qJycyQvj60OvRlH7mee8xAMjBDNRdMXlMwjAbMTDjkg=="
+    },
     "node_modules/grapheme-splitter": {
       "version": "1.0.4",
       "resolved": "https://registry.npmjs.org/grapheme-splitter/-/grapheme-splitter-1.0.4.tgz",
diff --git a/package.json b/package.json
index 0e43c41..2c52020 100644
--- a/package.json
+++ b/package.json
@@ -12,6 +12,7 @@
   "dependencies": {
     "discord.js": "^14.8.0",
     "fold-to-ascii": "^5.0.1",
+    "gpt-3-encoder": "^1.1.4",
     "openai": "^3.2.1"
   },
   "devDependencies": {
diff --git a/src/toOpenAIMessages.ts b/src/toOpenAIMessages.ts
index 5f97681..38c7cd1 100644
--- a/src/toOpenAIMessages.ts
+++ b/src/toOpenAIMessages.ts
@@ -3,6 +3,7 @@ import { Collection, Message as DiscordMessage } from "discord.js";
 import FoldToAscii from "fold-to-ascii";
 
 import config from "./config.json";
+import countTokens from "./tokenCounter";
 
 /**
  * Formats the message to use as a message content in OpenAI api
@@ -74,24 +75,29 @@ function getAuthorUsername(message: DiscordMessage): string {
 }
 
 /**
- * Converts the Collection of Discord Messages to array of OpenAI Messages
+ * Converts the Collection of Discord Messages to array of OpenAI Messages to send
  * @param messages the collection to convert
  * @returns the converted messages
  */
 export default function toOpenAIMessages(messages: Collection<string, DiscordMessage>): OpenAIMessage[] {
   const rvalue: OpenAIMessage[] = [];
+  let tokenCount = 0;
+
+  messages.sort((a, b) => b.createdTimestamp - a.createdTimestamp);
+
+  for (const message of messages.values()) {
+    const content = formatMessage(message);
+    // FIXME: tokens are not being counted properly (it's lower than it is) but it's enough for me for now.
+    tokenCount += countTokens(content);
+    if (tokenCount > 2048) break;
+    rvalue.push({
+      role: message.author.id == message.client.user.id ? "assistant" : "user",
+      content: content,
+      name: getAuthorUsername(message),
+    });
+  }
 
   rvalue.push({ role: "system", content: config.systemPrompt});
 
-  messages
-  .sort((a, b) => a.createdTimestamp - b.createdTimestamp)
-  .each(message => {
-    rvalue.push({
-      role: message.author.id == message.client.user.id ? "assistant" : "user",
-      content: formatMessage(message),
-      name: getAuthorUsername(message),
-    });
-  });
-
-  return rvalue;
+  return rvalue.reverse();
 }
diff --git a/src/tokenCounter.ts b/src/tokenCounter.ts
new file mode 100644
index 0000000..7141d0d
--- /dev/null
+++ b/src/tokenCounter.ts
@@ -0,0 +1,5 @@
+import { encode } from "gpt-3-encoder";
+
+export default function countTokens(text: string): number {
+    return encode(text).length;
+}