async fixes

app.py

@@ -2,12 +2,14 @@
 from __future__ import annotations
 import asyncio
 import datetime
-import pytz
 import logging
 import os
 from enum import Enum
 import json
 import uuid
+import threading
+
+import pytz
 from pydantic import BaseModel
 import gspread
 
@@ -23,6 +25,7 @@ import tiktoken
 
 from langchain.callbacks.streaming_aiter import AsyncIteratorCallbackHandler
 from langchain.callbacks.tracers.run_collector import RunCollectorCallbackHandler
+from langchain.callbacks.tracers.langchain import wait_for_all_tracers
 from langchain.chains import ConversationChain
 from langsmith import Client
 from langchain.chat_models import ChatAnthropic, ChatOpenAI
@@ -40,6 +43,8 @@ logging.basicConfig(format="%(asctime)s %(name)s %(levelname)s:%(message)s")
 LOG = logging.getLogger(__name__)
 LOG.setLevel(logging.INFO)
 
+thread_lock = threading.Lock()
+
 GPT_3_5_CONTEXT_LENGTH = 4096
 CLAUDE_2_CONTEXT_LENGTH = 100000  # need to use claude tokenizer
 
@@ -464,7 +469,7 @@ async def respond(
             state.chain.apredict(
                 input=chat_input,
                 callbacks=[callback, run_collector],
-            )
+            ),
         )
         state.history.append((chat_input, ""))
         run_id = None
@@ -474,7 +479,11 @@ async def respond(
             bot += tok
             state.history[-1] = (user, bot)
             yield state.history, state, None
-        await run
+        complete_response = await run
+        wait_for_all_tracers()
+        user, _ = state.history[-1]
+        state.history[-1] = (user, complete_response)
+        url_markdown = None
         if run_collector.traced_runs and run_id is None:
             run_id = run_collector.traced_runs[0].id
             LOG.info(f"RUNID: {run_id}")
@@ -495,7 +504,7 @@ async def respond(
                 ):
                     url_markdown += """\n
                     🙌 You have completed 10 exchanges with the chatbot."""
-                yield state.history, state, url_markdown
+        yield state.history, state, url_markdown
         LOG.info(f"""[{request.username}] ENDING CHAIN""")
         LOG.debug(f"History: {state.history}")
         LOG.debug(f"Memory: {state.chain.memory.json()}")
@@ -519,12 +528,13 @@ async def respond(
         gsheet_row = [[timestamp_string, *metadata_to_gsheet, langsmith_url]]
         LOG.info(f"Data to GSHEET: {gsheet_row}")
         try:
-            append_gsheet_rows(
-                sheet_id=GSHEET_ID,
-                sheet_name=TURNS_GSHEET_NAME,
-                rows=gsheet_row,
-                creds=GS_CREDS,
-            )
+            with thread_lock:
+                append_gsheet_rows(
+                    sheet_id=GSHEET_ID,
+                    sheet_name=TURNS_GSHEET_NAME,
+                    rows=gsheet_row,
+                    creds=GS_CREDS,
+                )
         except Exception as exc:
             LOG.error(f"Failed to log entry to Google Sheet. Row {gsheet_row}")
             LOG.error(exc)
@@ -657,4 +667,4 @@ with gr.Blocks(
     chat_submit_button.click(**clear_chatbot_messages_params)
     input_message.submit(**clear_chatbot_messages_params)
 
-demo.queue(max_size=99, concurrency_count=25, api_open=False).launch(auth=auth)
+demo.queue(max_size=25, concurrency_count=16, api_open=False).launch(auth=auth)