import os import traceback import asyncio from collections import defaultdict import discord import aiofiles from functools import partial from typing import List, Optional from pathlib import Path from datetime import date from langchain import OpenAI from gpt_index.readers import YoutubeTranscriptReader from gpt_index.readers.schema.base import Document from gpt_index import ( GPTSimpleVectorIndex, SimpleDirectoryReader, QuestionAnswerPrompt, BeautifulSoupWebReader, GPTListIndex, QueryMode, GPTTreeIndex, GoogleDocsReader, MockLLMPredictor, LLMPredictor, QueryConfig, PromptHelper, IndexStructType, OpenAIEmbedding, ) from gpt_index.readers.web import DEFAULT_WEBSITE_EXTRACTOR from gpt_index.composability import ComposableGraph from services.environment_service import EnvService, app_root_path SHORT_TO_LONG_CACHE = {} def get_and_query( user_id, index_storage, query, response_mode, nodes, llm_predictor, embed_model ): index: [GPTSimpleVectorIndex, ComposableGraph] = index_storage[ user_id ].get_index_or_throw() prompthelper = PromptHelper(4096, 500, 20) if isinstance(index, GPTTreeIndex): response = index.query( query, verbose=True, child_branch_factor=2, llm_predictor=llm_predictor, embed_model=embed_model, prompt_helper=prompthelper, ) else: response = index.query( query, response_mode=response_mode, verbose=True, llm_predictor=llm_predictor, embed_model=embed_model, similarity_top_k=nodes, prompt_helper=prompthelper, ) return response class IndexData: def __init__(self): self.queryable_index = None self.individual_indexes = [] # A safety check for the future def get_index_or_throw(self): if not self.queryable(): raise Exception( "An index access was attempted before an index was created. This is a programmer error, please report this to the maintainers." ) return self.queryable_index def queryable(self): return self.queryable_index is not None def has_indexes(self, user_id): try: return len(os.listdir(f"{app_root_path()}/indexes/{user_id}")) > 0 except Exception: return False def add_index(self, index, user_id, file_name): self.individual_indexes.append(index) self.queryable_index = index # Create a folder called "indexes/{USER_ID}" if it doesn't exist already Path(f"{app_root_path()}/indexes/{user_id}").mkdir(parents=True, exist_ok=True) # Save the index to file under the user id index.save_to_disk( app_root_path() / "indexes" / f"{str(user_id)}" / f"{file_name}_{date.today().month}_{date.today().day}.json" ) def reset_indexes(self, user_id): self.individual_indexes = [] self.queryable_index = None # Delete the user indexes try: # First, clear all the files inside it for file in os.listdir(f"{app_root_path()}/indexes/{user_id}"): os.remove(f"{app_root_path()}/indexes/{user_id}/{file}") except Exception: traceback.print_exc() class Index_handler: def __init__(self, bot, usage_service): self.bot = bot self.openai_key = os.getenv("OPENAI_TOKEN") self.index_storage = defaultdict(IndexData) self.loop = asyncio.get_running_loop() self.usage_service = usage_service self.qaprompt = QuestionAnswerPrompt( "Context information is below. The text '<|endofstatement|>' is used to separate chat entries and make it easier for you to understand the context\n" "---------------------\n" "{context_str}" "\n---------------------\n" "Never say '<|endofstatement|>'\n" "Given the context information and not prior knowledge, " "answer the question: {query_str}\n" ) # TODO We need to do predictions below for token usage. def index_file(self, file_path, embed_model) -> GPTSimpleVectorIndex: document = SimpleDirectoryReader(file_path).load_data() index = GPTSimpleVectorIndex(document, embed_model=embed_model) return index def index_gdoc(self, doc_id, embed_model) -> GPTSimpleVectorIndex: document = GoogleDocsReader().load_data(doc_id) index = GPTSimpleVectorIndex(document, embed_model=embed_model) return index def index_youtube_transcript(self, link, embed_model): documents = YoutubeTranscriptReader().load_data(ytlinks=[link]) index = GPTSimpleVectorIndex( documents, embed_model=embed_model, ) return index def index_load_file(self, file_path) -> [GPTSimpleVectorIndex, ComposableGraph]: if "composed_deep" in str(file_path): index = GPTTreeIndex.load_from_disk(file_path) else: index = GPTSimpleVectorIndex.load_from_disk(file_path) return index def index_discord(self, document, embed_model) -> GPTSimpleVectorIndex: index = GPTSimpleVectorIndex( document, embed_model=embed_model, ) return index def index_webpage(self, url, embed_model) -> GPTSimpleVectorIndex: documents = BeautifulSoupWebReader( website_extractor=DEFAULT_WEBSITE_EXTRACTOR ).load_data(urls=[url]) index = GPTSimpleVectorIndex(documents, embed_model=embed_model) return index def reset_indexes(self, user_id): self.index_storage[user_id].reset_indexes(user_id) async def set_file_index( self, ctx: discord.ApplicationContext, file: discord.Attachment, user_api_key ): if not user_api_key: os.environ["OPENAI_API_KEY"] = self.openai_key else: os.environ["OPENAI_API_KEY"] = user_api_key try: print(file.content_type) if file.content_type.startswith("text/plain"): suffix = ".txt" elif file.content_type.startswith("application/pdf"): suffix = ".pdf" # Allow for images too elif file.content_type.startswith("image/png"): suffix = ".png" elif file.content_type.startswith("image/"): suffix = ".jpg" elif "csv" in file.content_type: suffix = ".csv" elif "vnd." in file.content_type: suffix = ".pptx" # Catch all audio files and suffix with "mp3" elif file.content_type.startswith("audio/"): suffix = ".mp3" # Catch video files elif file.content_type.startswith("video/"): pass # No suffix change else: await ctx.respond( "Only accepts text, pdf, images, spreadheets, powerpoint, and audio/video files." ) return async with aiofiles.tempfile.TemporaryDirectory() as temp_path: async with aiofiles.tempfile.NamedTemporaryFile( suffix=suffix, dir=temp_path, delete=False ) as temp_file: await file.save(temp_file.name) embedding_model = OpenAIEmbedding() index = await self.loop.run_in_executor( None, partial(self.index_file, temp_path, embedding_model) ) await self.usage_service.update_usage( embedding_model.last_token_usage, embeddings=True ) file_name = file.filename self.index_storage[ctx.user.id].add_index(index, ctx.user.id, file_name) await ctx.respond("Index added to your indexes.") except Exception: await ctx.respond("Failed to set index") traceback.print_exc() async def set_link_index( self, ctx: discord.ApplicationContext, link: str, user_api_key ): if not user_api_key: os.environ["OPENAI_API_KEY"] = self.openai_key else: os.environ["OPENAI_API_KEY"] = user_api_key # TODO Link validation try: embedding_model = OpenAIEmbedding() # Check if the link contains youtube in it if "youtube" in link: index = await self.loop.run_in_executor( None, partial(self.index_youtube_transcript, link, embedding_model) ) else: index = await self.loop.run_in_executor( None, partial(self.index_webpage, link, embedding_model) ) await self.usage_service.update_usage( embedding_model.last_token_usage, embeddings=True ) # Make the url look nice, remove https, useless stuff, random characters file_name = ( link.replace("https://", "") .replace("http://", "") .replace("www.", "") .replace("/", "_") .replace("?", "_") .replace("&", "_") .replace("=", "_") .replace("-", "_") .replace(".", "_") ) self.index_storage[ctx.user.id].add_index(index, ctx.user.id, file_name) except Exception: await ctx.respond("Failed to set index") traceback.print_exc() await ctx.respond("Index set") async def set_discord_index( self, ctx: discord.ApplicationContext, channel: discord.TextChannel, user_api_key, ): if not user_api_key: os.environ["OPENAI_API_KEY"] = self.openai_key else: os.environ["OPENAI_API_KEY"] = user_api_key try: document = await self.load_data( channel_ids=[channel.id], limit=1000, oldest_first=False ) embedding_model = OpenAIEmbedding() index = await self.loop.run_in_executor( None, partial(self.index_discord, document, embedding_model) ) await self.usage_service.update_usage( embedding_model.last_token_usage, embeddings=True ) self.index_storage[ctx.user.id].add_index(index, ctx.user.id, channel.name) await ctx.respond("Index set") except Exception: await ctx.respond("Failed to set index") traceback.print_exc() async def load_index( self, ctx: discord.ApplicationContext, index, server, user_api_key ): if not user_api_key: os.environ["OPENAI_API_KEY"] = self.openai_key else: os.environ["OPENAI_API_KEY"] = user_api_key try: if server: index_file = EnvService.find_shared_file( f"indexes/{ctx.guild.id}/{index}" ) else: index_file = EnvService.find_shared_file( f"indexes/{ctx.user.id}/{index}" ) index = await self.loop.run_in_executor( None, partial(self.index_load_file, index_file) ) self.index_storage[ctx.user.id].queryable_index = index await ctx.respond("Loaded index") except Exception as e: await ctx.respond(e) async def compose_indexes(self, user_id, indexes, name, deep_compose): # Load all the indexes first index_objects = [] for _index in indexes: index_file = EnvService.find_shared_file(f"indexes/{user_id}/{_index}") index = await self.loop.run_in_executor( None, partial(self.index_load_file, index_file) ) index_objects.append(index) # For each index object, add its documents to a GPTTreeIndex if deep_compose: documents = [] for _index in index_objects: [ documents.append(_index.docstore.get_document(doc_id)) for doc_id in [docmeta for docmeta in _index.docstore.docs.keys()] if isinstance(_index.docstore.get_document(doc_id), Document) ] llm_predictor = LLMPredictor(llm=OpenAI(model_name="text-davinci-003")) embedding_model = OpenAIEmbedding() tree_index = GPTTreeIndex( documents=documents, llm_predictor=llm_predictor, embed_model=embedding_model, ) await self.usage_service.update_usage(llm_predictor.last_token_usage) await self.usage_service.update_usage( embedding_model.last_token_usage, embeddings=True ) # Now we have a list of tree indexes, we can compose them if not name: name = ( f"composed_deep_index_{date.today().month}_{date.today().day}.json" ) # Save the composed index tree_index.save_to_disk(f"indexes/{user_id}/{name}.json") self.index_storage[user_id].queryable_index = tree_index else: documents = [] for _index in index_objects: [ documents.append(_index.docstore.get_document(doc_id)) for doc_id in [docmeta for docmeta in _index.docstore.docs.keys()] if isinstance(_index.docstore.get_document(doc_id), Document) ] embedding_model = OpenAIEmbedding() # Add everything into a simple vector index simple_index = GPTSimpleVectorIndex( documents=documents, embed_model=embedding_model ) await self.usage_service.update_usage( embedding_model.last_token_usage, embeddings=True ) if not name: name = f"composed_index_{date.today().month}_{date.today().day}.json" # Save the composed index simple_index.save_to_disk(f"indexes/{user_id}/{name}.json") self.index_storage[user_id].queryable_index = simple_index async def backup_discord(self, ctx: discord.ApplicationContext, user_api_key): if not user_api_key: os.environ["OPENAI_API_KEY"] = self.openai_key else: os.environ["OPENAI_API_KEY"] = user_api_key try: channel_ids: List[int] = [] for c in ctx.guild.text_channels: channel_ids.append(c.id) document = await self.load_data( channel_ids=channel_ids, limit=3000, oldest_first=False ) embedding_model = OpenAIEmbedding() index = await self.loop.run_in_executor( None, partial(self.index_discord, document, embedding_model) ) await self.usage_service.update_usage( embedding_model.last_token_usage, embeddings=True ) Path(app_root_path() / "indexes" / str(ctx.guild.id)).mkdir( parents=True, exist_ok=True ) index.save_to_disk( app_root_path() / "indexes" / str(ctx.guild.id) / f"{ctx.guild.name.replace(' ', '-')}_{date.today().month}_{date.today().day}.json" ) await ctx.respond("Backup saved") except Exception: await ctx.respond("Failed to save backup") traceback.print_exc() async def query( self, ctx: discord.ApplicationContext, query: str, response_mode, nodes, user_api_key, ): if not user_api_key: os.environ["OPENAI_API_KEY"] = self.openai_key else: os.environ["OPENAI_API_KEY"] = user_api_key try: llm_predictor = LLMPredictor(llm=OpenAI(model_name="text-davinci-003")) embedding_model = OpenAIEmbedding() embedding_model.last_token_usage = 0 response = await self.loop.run_in_executor( None, partial( get_and_query, ctx.user.id, self.index_storage, query, response_mode, nodes, llm_predictor, embedding_model, ), ) print("The last token usage was ", llm_predictor.last_token_usage) await self.usage_service.update_usage(llm_predictor.last_token_usage) await self.usage_service.update_usage( embedding_model.last_token_usage, embeddings=True ) await ctx.respond( f"**Query:**\n\n{query.strip()}\n\n**Query response:**\n\n{response.response.strip()}" ) except Exception: traceback.print_exc() await ctx.respond( "Failed to send query. You may not have an index set, load an index with /index load", delete_after=10, ) # Extracted functions from DiscordReader async def read_channel( self, channel_id: int, limit: Optional[int], oldest_first: bool ) -> str: """Async read channel.""" messages: List[discord.Message] = [] try: channel = self.bot.get_channel(channel_id) print(f"Added {channel.name} from {channel.guild.name}") # only work for text channels for now if not isinstance(channel, discord.TextChannel): raise ValueError( f"Channel {channel_id} is not a text channel. " "Only text channels are supported for now." ) # thread_dict maps thread_id to thread thread_dict = {} for thread in channel.threads: thread_dict[thread.id] = thread async for msg in channel.history(limit=limit, oldest_first=oldest_first): if msg.author.bot: pass else: messages.append(msg) if msg.id in thread_dict: thread = thread_dict[msg.id] async for thread_msg in thread.history( limit=limit, oldest_first=oldest_first ): messages.append(thread_msg) except Exception as e: print("Encountered error: " + str(e)) channel = self.bot.get_channel(channel_id) msg_txt_list = [ f"user:{m.author.display_name}, content:{m.content}" for m in messages ] return ("<|endofstatement|>\n\n".join(msg_txt_list), channel.name) async def load_data( self, channel_ids: List[int], limit: Optional[int] = None, oldest_first: bool = True, ) -> List[Document]: """Load data from the input directory. Args: channel_ids (List[int]): List of channel ids to read. limit (Optional[int]): Maximum number of messages to read. oldest_first (bool): Whether to read oldest messages first. Defaults to `True`. Returns: List[Document]: List of documents. """ results: List[Document] = [] for channel_id in channel_ids: if not isinstance(channel_id, int): raise ValueError( f"Channel id {channel_id} must be an integer, " f"not {type(channel_id)}." ) (channel_content, channel_name) = await self.read_channel( channel_id, limit=limit, oldest_first=oldest_first ) results.append( Document(channel_content, extra_info={"channel_name": channel_name}) ) return results async def compose(self, ctx: discord.ApplicationContext, name, user_api_key): # Send the ComposeModal if not user_api_key: os.environ["OPENAI_API_KEY"] = self.openai_key else: os.environ["OPENAI_API_KEY"] = user_api_key if not self.index_storage[ctx.user.id].has_indexes(ctx.user.id): await ctx.respond("You must load at least one indexes before composing") return await ctx.respond( "Select the index(es) to compose. You can compose multiple indexes together, you can also Deep Compose a single index.", view=ComposeModal(self, ctx.user.id, name), ephemeral=True, ) class ComposeModal(discord.ui.View): def __init__(self, index_cog, user_id, name=None, deep=None) -> None: super().__init__() # Get the argument named "user_key_db" and save it as USER_KEY_DB self.index_cog = index_cog self.user_id = user_id self.deep = deep # Get all the indexes for the user self.indexes = [ file for file in os.listdir( EnvService.find_shared_file(f"indexes/{str(user_id)}/") ) ] # Map everything into the short to long cache for index in self.indexes: SHORT_TO_LONG_CACHE[index[:99]] = index # A text entry field for the name of the composed index self.name = name # A discord UI select menu with all the indexes. Limited to 25 entries. For the label field in the SelectOption, # cut it off at 100 characters to prevent the message from being too long self.index_select = discord.ui.Select( placeholder="Select index(es) to compose", options=[ discord.SelectOption(label=str(index)[:99], value=index[:99]) for index in self.indexes ][0:25], max_values=len(self.indexes) if len(self.indexes) < 25 else 25, min_values=1, ) # Add the select menu to the modal self.add_item(self.index_select) # If we have more than 25 entries, add more Select fields as neccessary self.extra_index_selects = [] if len(self.indexes) > 25: for i in range(25, len(self.indexes), 25): self.extra_index_selects.append( discord.ui.Select( placeholder="Select index(es) to compose", options=[ discord.SelectOption(label=index[:99], value=index[:99]) for index in self.indexes ][i : i + 25], max_values=len(self.indexes[i : i + 25]), min_values=1, ) ) self.add_item(self.extra_index_selects[-1]) # Add an input field for "Deep", a "yes" or "no" option, default no self.deep_select = discord.ui.Select( placeholder="Deep Compose", options=[ discord.SelectOption(label="Yes", value="yes"), discord.SelectOption(label="No", value="no"), ], max_values=1, min_values=1, ) self.add_item(self.deep_select) # Add a button to the modal called "Compose" self.add_item( discord.ui.Button( label="Compose", style=discord.ButtonStyle.green, custom_id="compose" ) ) # The callback for the button async def interaction_check(self, interaction: discord.Interaction) -> bool: # Check that the interaction was for custom_id "compose" if interaction.data["custom_id"] == "compose": # Check that the user selected at least one index # The total list of indexes is the union of the values of all the select menus indexes = self.index_select.values + [ select.values[0] for select in self.extra_index_selects ] # Remap them from the SHORT_TO_LONG_CACHE indexes = [SHORT_TO_LONG_CACHE[index] for index in indexes] if len(indexes) < 1: await interaction.response.send_message( "You must select at least 1 index", ephemeral=True ) else: composing_message = await interaction.response.send_message( "Composing indexes, this may take a long time...", ephemeral=True, delete_after=120, ) # Compose the indexes await self.index_cog.compose_indexes( self.user_id, indexes, self.name, False if not self.deep_select.values or self.deep_select.values[0] == "no" else True, ) await interaction.followup.send( "Composed indexes", ephemeral=True, delete_after=10 ) try: await composing_message.delete() except: pass else: await interaction.response.defer(ephemeral=True)