index instabilit

gpt3discord.py

@@ -34,7 +34,7 @@
 from models.openai_model import Model
 
 
-__version__ = "12.1.6"
+__version__ = "12.1.7"
 
 
 PID_FILE = Path("bot.pid")

models/index_model.py

@@ -9,6 +9,7 @@
 import aiohttp
 import discord
 import aiofiles
+import httpx
 import openai
 import tiktoken
 from functools import partial
@@ -84,7 +85,6 @@
 from models.openai_model import Models
 from models.check_model import UrlCheck
 from services.environment_service import EnvService
-
 SHORT_TO_LONG_CACHE = {}
 MAX_DEEP_COMPOSE_PRICE = EnvService.get_max_deep_compose_price()
 EpubReader = download_loader("EpubReader")
@@ -101,11 +101,21 @@
     text_splitter=TokenTextSplitter(chunk_size=1024, chunk_overlap=20)
 )
 callback_manager = CallbackManager([token_counter])
-service_context = ServiceContext.from_defaults(
+service_context_no_llm = ServiceContext.from_defaults(
     embed_model=embedding_model,
     callback_manager=callback_manager,
     node_parser=node_parser,
 )
+timeout = httpx.Timeout(1, read=1, write=1, connect=1)
+
+def get_service_context_with_llm(llm):
+    service_context = ServiceContext.from_defaults(
+        embed_model=embedding_model,
+        callback_manager=callback_manager,
+        node_parser=node_parser,
+        llm=llm,
+    )
+    return service_context
 
 
 def dummy_tool(**kwargs):
@@ -400,21 +410,25 @@ async def index_chat_file(self, message: discord.Message, file: discord.Attachme
                     # Assert that the filename is < 100 characters, if it is greater, truncate to the first 100 characters and keep the original ending
                     if len(filename) > 100:
                         filename = filename[:100] + filename[-4:]
+                    openai.log = "debug"
 
+                    print("Indexing")
                     index: VectorStoreIndex = await self.loop.run_in_executor(
                         None,
                         partial(
                             self.index_file,
                             Path(temp_file.name),
-                            service_context,
+                            get_service_context_with_llm(self.index_chat_chains[message.channel.id].llm),
                             suffix,
                         ),
                     )
+                    print("Done Indexing")
 
                     summary = await index.as_query_engine(
                         similarity_top_k=10,
-                        child_branch_factor=6,
+                        child_branch_factor=3,
                         response_mode="tree_summarize",
+                        service_context=get_service_context_with_llm(self.index_chat_chains[message.channel.id].llm)
                     ).aquery(
                         f"What is a summary or general idea of this data? Be detailed in your summary (e.g "
                         f"extract key names, etc) but not too verbose. Your summary should be under a hundred words. "
@@ -427,7 +441,7 @@ async def index_chat_file(self, message: discord.Message, file: discord.Attachme
                         f"is no available data if there are no available tools that are relevant."
                     )
 
-                    engine = self.get_query_engine(index, message, summary)
+                    engine = self.get_query_engine(index, self.index_chat_chains[message.channel.id].llm)
 
                     # Get rid of all special characters in the filename
                     filename = "".join(
@@ -485,8 +499,7 @@ async def start_index_chat(self, ctx, model):
         preparation_message = await ctx.channel.send(
             embed=EmbedStatics.get_index_chat_preparation_message()
         )
-
-        llm = ChatOpenAI(model=model, temperature=0)
+        llm = ChatOpenAI(model=model, temperature=0, max_retries=2)
         llm_predictor = LLMPredictor(llm=ChatOpenAI(temperature=0, model_name=model))
 
         max_token_limit = 29000 if "gpt-4" in model else 7500
@@ -508,7 +521,7 @@ async def start_index_chat(self, ctx, model):
                 "to the data at the link by the time you respond. When using tools, the input should be "
                 "clearly created based on the request of the user. For example, if a user uploads an invoice "
                 "and asks how many usage hours of X was present in the invoice, a good query is 'X hours'. "
-                "Avoid using single word queries unless the request is very simple. You can query multiple times to break down complex requests and retrieve more information."
+                "Avoid using single word queries unless the request is very simple. You can query multiple times to break down complex requests and retrieve more information. When calling functions, no special characters are allowed in the function name, keep that in mind."
             ),
         }
 
@@ -786,7 +799,7 @@ async def set_file_index(
                         partial(
                             self.index_file,
                             Path(temp_file.name),
-                            service_context,
+                            service_context_no_llm,
                             suffix,
                         ),
                     )
@@ -858,7 +871,7 @@ async def set_link_index_recurse(
                 functools.partial(
                     GPTVectorStoreIndex,
                     documents=documents,
-                    service_context=service_context,
+                    service_context=service_context_no_llm,
                     use_async=True,
                 ),
             )
@@ -902,16 +915,16 @@ async def set_link_index_recurse(
 
         await response.edit(embed=EmbedStatics.get_index_set_success_embed(price))
 
-    def get_query_engine(self, index, message, summary):
+    def get_query_engine(self, index, llm):
         retriever = VectorIndexRetriever(
-            index=index, similarity_top_k=6, service_context=service_context
+            index=index, similarity_top_k=6, service_context=get_service_context_with_llm(llm)
         )
 
         response_synthesizer = get_response_synthesizer(
             response_mode=ResponseMode.COMPACT_ACCUMULATE,
             use_async=True,
             refine_template=TEXT_QA_SYSTEM_PROMPT,
-            service_context=service_context,
+            service_context=get_service_context_with_llm(llm),
             verbose=True,
         )
 
@@ -924,15 +937,17 @@ def get_query_engine(self, index, message, summary):
     async def index_link(self, link, summarize=False, index_chat_ctx=None):
         try:
             if await UrlCheck.check_youtube_link(link):
+                print("Indexing youtube transcript")
                 index = await self.loop.run_in_executor(
-                    None, partial(self.index_youtube_transcript, link, service_context)
+                    None, partial(self.index_youtube_transcript, link, service_context_no_llm)
                 )
+                print("Indexed youtube transcript")
             elif "github" in link:
                 index = await self.loop.run_in_executor(
-                    None, partial(self.index_github_repository, link, service_context)
+                    None, partial(self.index_github_repository, link, service_context_no_llm)
                 )
             else:
-                index = await self.index_webpage(link, service_context)
+                index = await self.index_webpage(link, service_context_no_llm)
         except Exception as e:
             if index_chat_ctx:
                 await index_chat_ctx.reply(
@@ -945,13 +960,16 @@ async def index_link(self, link, summarize=False, index_chat_ctx=None):
         summary = None
         if index_chat_ctx:
             try:
+                print("Getting transcript summary")
                 summary = await index.as_query_engine(
-                    response_mode="tree_summarize"
+                    response_mode="tree_summarize",
+                    service_context=get_service_context_with_llm(self.index_chat_chains[index_chat_ctx.channel.id].llm)
                 ).aquery(
                     "What is a summary or general idea of this document? Be detailed in your summary but not too verbose. Your summary should be under 50 words. This summary will be used in a vector index to retrieve information about certain data. So, at a high level, the summary should describe the document in such a way that a retriever would know to select it when asked questions about it. The link was {link}. Include the an easy identifier derived from the link at the end of the summary."
                 )
+                print("Got transcript summary")
 
-                engine = self.get_query_engine(index, index_chat_ctx, summary)
+                engine = self.get_query_engine(index, self.index_chat_chains[index_chat_ctx.channel.id].llm)
 
                 # Get rid of all special characters in the link, replace periods with _
                 link_cleaned = "".join(
@@ -1078,7 +1096,7 @@ async def set_discord_index(
                 channel_ids=[channel.id], limit=message_limit, oldest_first=False
             )
             index = await self.loop.run_in_executor(
-                None, partial(self.index_discord, document, service_context)
+                None, partial(self.index_discord, document, service_context_no_llm)
             )
             try:
                 price = await self.usage_service.get_price(
@@ -1255,7 +1273,7 @@ async def compose_indexes(self, user_id, indexes, name, deep_compose):
                 partial(
                     GPTVectorStoreIndex.from_documents,
                     documents=documents,
-                    service_context=service_context,
+                    service_context=service_context_no_llm,
                     use_async=True,
                 ),
             )
@@ -1300,7 +1318,7 @@ async def backup_discord(
             )
 
             index = await self.loop.run_in_executor(
-                None, partial(self.index_discord, document, service_context)
+                None, partial(self.index_discord, document, service_context_no_llm)
             )
             await self.usage_service.update_usage(
                 token_counter.total_embedding_token_count, "embedding"
@@ -1362,7 +1380,7 @@ async def query(
                     response_mode,
                     nodes,
                     child_branch_factor,
-                    service_context=service_context,
+                    service_context=service_context_no_llm,
                     multistep=llm_predictor if multistep else None,
                 ),
             )

pyproject.toml

@@ -19,7 +19,7 @@ classifiers = [
 ]
 dependencies = [
 "Pillow==9.3.0",
-"openai==1.1.0",
+"openai==1.2.0",
 "yt-dlp==2023.3.4",
 "ffmpeg==1.4",
 "beautifulsoup4==4.11.2",
@@ -40,7 +40,7 @@ dependencies = [
 "sentencepiece==0.1.99",
 "protobuf==3.20.2",
 "python-pptx==0.6.21",
-"langchain==0.0.332",
+"langchain==0.0.333",
 "unidecode==1.3.6",
 "tqdm==4.64.1",
 "docx2txt==0.8",

requirements.txt

@@ -1,5 +1,5 @@
 Pillow==9.3.0
-openai==1.1.0
+openai==1.2.0
 yt-dlp==2023.3.4
 ffmpeg==1.4
 py-cord==2.4.1
@@ -21,7 +21,7 @@ sentencepiece==0.1.99
 protobuf==3.20.2
 python-pptx==0.6.21
 sentence-transformers==2.2.2
-langchain==0.0.332
+langchain==0.0.333
 openai-whisper
 unidecode==1.3.6
 tqdm==4.64.1

requirements_base.txt

@@ -1,5 +1,5 @@
 Pillow==9.3.0
-openai==1.1.0
+openai==1.2.0
 yt-dlp==2023.3.4
 ffmpeg==1.4
 py-cord==2.4.1
@@ -20,7 +20,7 @@ youtube_transcript_api==0.5.0
 sentencepiece==0.1.99
 protobuf==3.20.2
 python-pptx==0.6.21
-langchain==0.0.332
+langchain==0.0.333
 unidecode==1.3.6
 tqdm==4.64.1
 docx2txt==0.8