Merge pull request #1 from VladVP/main

Add initial system prompt to create behavior identical with and/or equivalent to ChatGPT on the web.

Author: egbaydarov

.env.example

@@ -3,4 +3,9 @@ DISCORD_BOT_TOKEN=x
 DISCORD_CLIENT_ID=x
 
 ALLOWED_SERVER_IDS=1
-SERVER_TO_MODERATION_CHANNEL=1:1
+
+OPENAI_API_URL=https://api.openai.com/v1/chat/completions
+OPENAI_MODEL=gpt-3.5-turbo
+
+SYSTEM_MESSAGE="You are ChatGPT, a large language model trained by OpenAI. Answer as concisely as possible. Knowledge cutoff: {knowledge_cutoff} Current date: {current_date}"
+KNOWLEDGE_CUTOFF="2021-09"

README.md

@@ -16,9 +16,7 @@ Thank you!
 ---
 # GPT Discord Bot
 
-Example Discord bot written in Python that uses the [completions API](https://beta.openai.com/docs/api-reference/completions) to have conversations with the `text-davinci-003` model, and the [moderations API](https://beta.openai.com/docs/api-reference/moderations) to filter the messages.
-
-**THIS IS NOT CHATGPT.**
+Example Discord bot written in Python that uses the [completions API](https://beta.openai.com/docs/api-reference/completions) to have conversations with the `gpt-3.5-turbo` or `gpt-4` models.
 
 This bot uses the [OpenAI Python Library](https://github.com/openai/openai-python) and [discord.py](https://discordpy.readthedocs.io/).
 
@@ -53,9 +51,9 @@ This bot uses the [OpenAI Python Library](https://github.com/openai/openai-pytho
 
 # Optional configuration
 
-1. If you want moderation messages, create and copy the channel id for each server that you want the moderation messages to send to in `SERVER_TO_MODERATION_CHANNEL`. This should be of the format: `server_id:channel_id,server_id_2:channel_id_2`
-1. If you want to change the personality of the bot, go to `src/config.yaml` and edit the instructions
-1. If you want to change the moderation settings for which messages get flagged or blocked, edit the values in `src/constants.py`. A lower value means less chance of it triggering.
+- If you want to change the model used, you can do so in `OPENAI_MODEL`. Currently only `gpt-3.5-turbo` and `gpt-4` work with the present codebase.
+
+- If you want to change the behavior/personality of the bot, change the system prompt in `SYSTEM_MESSAGE`, with optional variables enclosed in `{`curly braces`}`. Currently the only variables available are `current_date` and `knowledge_cutoff`, with the latter being equivalent to the environment variable of the same name. The former is always in ISO 8601 format.
 
 # FAQ
 

src/constants.py

@@ -11,6 +11,9 @@
 OPENAI_API_URL = os.environ['OPENAI_API_URL']
 OPENAI_MODEL = os.environ['OPENAI_MODEL']
 
+SYSTEM_MESSAGE = os.environ["SYSTEM_MESSAGE"]
+KNOWLEDGE_CUTOFF = os.environ["KNOWLEDGE_CUTOFF"]
+
 ALLOWED_SERVER_IDS: List[int] = []
 server_ids = os.environ["ALLOWED_SERVER_IDS"].split(",")
 for s in server_ids:
@@ -36,5 +39,5 @@
 ACTIVATE_THREAD_PREFX = "💬✅"
 INACTIVATE_THREAD_PREFIX = "💬❌"
 MAX_CHARS_PER_REPLY_MSG = (
-    1500  # discord has a 2k limit, we just break message into 1.5k
+    2000  # discord has a 2k limit
 )

src/main.py

@@ -1,3 +1,4 @@
+import datetime
 import discord
 from discord import Message as DiscordMessage
 import logging
@@ -8,6 +9,8 @@
     ACTIVATE_THREAD_PREFX,
     MAX_THREAD_MESSAGES,
     SECONDS_DELAY_RECEIVING_MSG,
+    SYSTEM_MESSAGE,
+    KNOWLEDGE_CUTOFF
 )
 import asyncio
 from src.utils import (
@@ -78,8 +81,17 @@ async def chat_command(int: discord.Interaction, message: str):
             auto_archive_duration=60,
         )
         async with thread.typing():
+            # prepare the initial system message
+            current_date = datetime.datetime.now().strftime("%Y-%m-%d")
+            system_message = SYSTEM_MESSAGE.format(
+                knowledge_cutoff=KNOWLEDGE_CUTOFF,
+                current_date=current_date
+            )
             # fetch completion
-            messages = [Message(user='system', text=message)]
+            messages = [
+                Message(user='system', text=system_message),
+                Message(user='user', text=message)
+            ]
             response_data = await generate_completion_response(
                 messages=messages,
             )
@@ -145,13 +157,21 @@ async def on_message(message: DiscordMessage):
             f"Thread message to process - {message.author}: {message.content[:50]} - {thread.name} {thread.jump_url}"
         )
 
+        # prepare the initial system message
+        current_date = datetime.datetime.now().strftime("%Y-%m-%d")
+        system_message = SYSTEM_MESSAGE.format(
+            knowledge_cutoff=KNOWLEDGE_CUTOFF,
+            current_date=current_date
+        )
+
         channel_messages = [
             discord_message_to_message(
                 message=message,
                 bot_name=client.user)
             async for message in thread.history(limit=MAX_THREAD_MESSAGES)
         ]
         channel_messages = [x for x in channel_messages if x is not None]
+        channel_messages.append(Message(user='system', text=system_message))
         channel_messages.reverse()
 
         # generate the response

src/utils.py

@@ -26,30 +26,65 @@ def discord_message_to_message(
         logger.info(
             f"field.name - {field.name}"
         )
-        return Message(user="system", text=field.value)
+        return Message(user="user", text=field.value)
     else:
         if message.content:
             user_name = "assistant" if message.author == bot_name else "user"
             return Message(user=user_name, text=message.content)
     return None
 
 
-def split_into_shorter_messages(message: str) -> List[str]:
-    indices_object = re.finditer(
-        pattern='```',
-        string=message)
-
-    indices = [index.start() for index in indices_object]
-    indices[1::2] = [x + 4 for x in indices[1::2]]
-    indices.insert(0, 0)
-    indices.append(len(message))
-
-    result = []
-    for i in range(1, len(indices)):
-        result.append(message[indices[i-1]:indices[i]])
-
-    return result
-
+def split_into_shorter_messages(text, limit=MAX_CHARS_PER_REPLY_MSG, code_block='```'):
+    def split_at_boundary(s, boundary):
+        parts = s.split(boundary)
+        result = []
+        for i, part in enumerate(parts):
+            if i % 2 == 1:
+                result.extend(split_code_block(part))
+            else:
+                result += split_substring(part)
+        return result
+
+    def split_substring(s):
+        if len(s) <= limit:
+            return [s]
+        for boundary in ('\n', ' '):
+            if boundary in s:
+                break
+        else:
+            return [s[:limit]] + split_substring(s[limit:])
+
+        pieces = s.split(boundary)
+        result = []
+        current_part = pieces[0]
+        for piece in pieces[1:]:
+            if len(current_part) + len(boundary) + len(piece) > limit:
+                result.append(current_part)
+                current_part = piece
+            else:
+                current_part += boundary + piece
+        result.append(current_part)
+        return result
+
+    def split_code_block(s):
+        if len(code_block + s + code_block) <= limit:
+            return [code_block + s + code_block]
+        else:
+            lines = s.split('\n')
+            result = [code_block]
+            for line in lines:
+                if len(result[-1] + '\n' + line) > limit:
+                    result[-1] += code_block
+                    result.append(code_block + line)
+                else:
+                    result[-1] += '\n' + line
+            result[-1] += code_block
+            return result
+
+    if code_block in text:
+        return split_at_boundary(text, code_block)
+    else:
+        return split_substring(text)
 
 def is_last_message_stale(
     interaction_message: DiscordMessage, last_message: DiscordMessage, bot_id: str