Stream chat

chat.py

@@ -4,6 +4,7 @@
 from render_body import render_body
 from render_my_conversations import render_my_conversations
 import streamlit as st
+from firestore_utils import clear_user_history
 
 from dotenv import load_dotenv
 
@@ -29,6 +30,7 @@ def load_and_store_conversation(st, cid: str):
 def controller():
     # TODO: display useful total cost in $
     st.session_state["total_cost"] = 0.0
+    st.session_state["conversation_expanded"] = True
 
     # set model in session if specified in params
     model_from_param = get_key_from_params(st, "model")
@@ -88,9 +90,36 @@ def render_history_menu(sidebar):
     render_my_conversations(st, sidebar)
 
 
+def render_profile(sidebar):
+    user_info = st.session_state.get("user_info")
+    if not user_info:
+        return
+
+    status = f"Signed in as {user_info.get('email')}"
+    with sidebar.expander(status):
+        st.image(user_info.get("picture"), width=50)
+        signout = st.button("Sign out", key="button_signout", type="primary")
+        if signout:
+            st.session_state.clear()
+            st.experimental_rerun()
+        st.write(
+            "While it's useful to resume past conversations, sometimes you may want to clear your chat history."
+        )
+        placeholder = st.empty()
+        with placeholder:
+            clear_history = st.button(
+                "Clear History", key="button_clear_history", type="primary"
+            )
+        if clear_history:
+            clear_user_history(user_info["id"])
+            placeholder.info("Chat history cleared", icon="✅")
+            st.snow()
+
+
 def render_sidebar(sidebar):
     render_new_chat(sidebar)
     render_auth(st)
+    render_profile(sidebar)
     sidebar.divider()
     render_history_menu(sidebar)
 

content/overview.md

@@ -4,12 +4,11 @@ Thanks for trying this out. It's super alpha, but each person can log in via Goo
 
 Limitations:
 
-- you have to log in again if you reload the page (need to work around streamlit later)
-- openai api gives rate limit errors sometimes, causing crashes
-- gpt4 feels slower than ai.com naturally
-- can't delete personal conversations yet. don't put anything sensitive there.
+- **you have to log in again if you reload** the page (streamlit's limitation; work around later)
+- openai api response can be slow, with rate limit errors sometimes (chat.openai.com has advantage)
+- can only delete all chat history, not individual conversation yet (contribution welcome!)
 
 How to help:
 
 - report bugs to Hien or myself.
-- run your code locally https://github.com/CoderPush/chatlit by getting your own keys at .env
+- run your code locally https://github.com/CoderPush/chatlit by asking me for .env file

render_body.py

@@ -1,11 +1,10 @@
 from render_conversation import render_conversation
-from render_chat_form import render_chat_form
+from render_chat_form import render_chat_form, render_chat_stream
 from firestore_utils import get_firestore_db
-from utils import get_cid_from_params
-from firestore_utils import clear_user_history
+from utils import get_cid_from_params, get_expander_text
 
 
-def load_conversation(st):
+def load_conversation_from_db(st):
     cid = get_cid_from_params(st)
     if cid:
         db = get_firestore_db()
@@ -14,45 +13,24 @@ def load_conversation(st):
     return {}
 
 
-def get_expander_text(st):
-    user = st.session_state.get("user_info", {}).get("name", None)
-    model = st.session_state.get("model")
-    if user:
-        text = f"### {model} with {user}"
-    else:
-        text = f"### {model}"
-    return text
+def load_conversation_from_session_state(st):
+    return st.session_state.get("conversation", {})
 
 
 def render_body(st):
-    with st.expander(get_expander_text(st)):
-        user_info = st.session_state.get("user_info")
-        if user_info:
-            st.write(f"Signed in as {user_info.get('email')}")
-            st.image(user_info.get("picture"), width=50)
-            signout = st.button("Sign out", key="button_signout", type="primary")
-            if signout:
-                st.session_state.clear()
-                st.experimental_rerun()
-            st.write(
-                "While it's useful to resume past conversations, sometimes you may want to clear your chat history."
-            )
-            placeholder = st.empty()
-            with placeholder:
-                clear_history = st.button(
-                    "Clear History", key="button_clear_history", type="primary"
-                )
-            if clear_history:
-                clear_user_history(user_info["id"])
-                placeholder.info("Chat history cleared", icon="✅")
-                st.snow()
-
-    conversation = load_conversation(st)
-    if conversation:
-        render_conversation(st, conversation)
+    messages_holder = st.expander(
+        get_expander_text(st), expanded=st.session_state["conversation_expanded"]
+    )
+    with messages_holder:
+        # load_conversation from session_state
+        conversation = load_conversation_from_session_state(st)
+        if conversation:
+            render_conversation(st, conversation)
 
     if st.session_state.get("user_info"):
-        render_chat_form(st)
+        with st.container():
+            # render_chat_form(st)
+            render_chat_stream(st)
     else:
         # load homepage.md into a string
         with open("content/overview.md", "r") as f:

render_chat_form.py

@@ -1,13 +1,21 @@
 import openai
 from firestore_utils import firestore_save
 from utils import generate_conversation_title, get_oauth_uid
+import time
 
 
-def load_messages(st):
+def extract_messages(st):
+    default = [
+        {
+            "role": "system",
+            "content": "You are a helpful assistant. Please use concise language to save bandwidth and token usage. Avoid 'AI language model' disclaimer whenever possible.",
+        }
+    ]
+
     conversation = st.session_state.get("conversation", {})
-    default = [{"role": "system", "content": "You are a helpful assistant."}]
+    messages = conversation.get("messages", default)
 
-    return conversation.get("messages", default)
+    return messages
 
 
 def get_content(st, response):
@@ -23,7 +31,7 @@ def get_content(st, response):
 
 def generate_response(st, prompt):
     model = st.session_state["model"]
-    messages = load_messages(st)
+    messages = extract_messages(st)
     messages.append({"role": "user", "content": prompt})
 
     print("openai.ChatCompletion.create with")
@@ -43,7 +51,7 @@ def generate_response(st, prompt):
     return messages, usage
 
 
-def save_to_firestore(st, messages, usage):
+def save_to_firestore(st, messages, usage=None):
     model = st.session_state["model"]
     if len(messages) > 0:
         conversation = st.session_state.get("conversation", {})
@@ -69,6 +77,7 @@ def save_to_firestore(st, messages, usage):
         return new_conversation
 
 
+# non-streaming version, with usage
 def render_chat_form(st):
     name = st.session_state.get("user_info", {}).get("name", "You")
     model = st.session_state["model"]
@@ -85,3 +94,83 @@ def render_chat_form(st):
         if new_conversation is not None:
             st.experimental_set_query_params(cid=new_conversation.id)
         st.experimental_rerun()
+
+
+# see sample-stream.json to know how to parse it
+def generate_stream(st, holder, user_input):
+    model = st.session_state["model"]
+    messages = extract_messages(st)
+    messages.append({"role": "user", "content": user_input})
+
+    print("openai.ChatCompletion.create with", model, messages)
+    completion = openai.ChatCompletion.create(
+        model=model,
+        messages=messages,
+        stream=True,
+    )
+
+    # first chunk should be
+    # {
+    #     "choices": [
+    #     {
+    #         "delta": {
+    #         "role": "assistant"
+    #         },
+    #         "finish_reason": null,
+    #         "index": 0
+    #     }
+    #     ],
+    #     "created": 1684389483,
+    #     "id": "chatcmpl-7HQwF5QPvTrDtYPOvBZbzFfDb9tcI",
+    #     "model": "gpt-3.5-turbo-0301",
+    #     "object": "chat.completion.chunk"
+    # }
+
+    # middle chunks are content:
+    with holder.container():
+        content = ""
+        for chunk in completion:
+            delta = chunk["choices"][0]["delta"]
+            if "content" in delta:
+                content += delta["content"]
+                holder.markdown(content)
+
+    # last chunk should be
+    # {
+    #     "choices": [
+    #     {
+    #         "delta": {},
+    #         "finish_reason": "stop",
+    #         "index": 0
+    #     }
+    #     ],
+    #     "created": 1684389483,
+    #     "id": "chatcmpl-7HQwF5QPvTrDtYPOvBZbzFfDb9tcI",
+    #     "model": "gpt-3.5-turbo-0301",
+    #     "object": "chat.completion.chunk"
+    # }
+
+    messages.append({"role": "assistant", "content": content})
+
+    # No usage info in stream mode yet
+    # https://community.openai.com/t/usage-info-in-api-responses/18862
+
+    return messages
+
+
+def render_chat_stream(st):
+    with st.form(key="chat_prompt", clear_on_submit=True):
+        stream_holder = st.empty()
+        user_input = st.text_area(
+            f"You:", key="text_area_stream", label_visibility="collapsed"
+        )
+        submit_button = st.form_submit_button(label="Send")
+
+    if submit_button and user_input:
+        st.session_state["conversation_expanded"] = False
+        messages = generate_stream(st, stream_holder, user_input)
+        # print("messages: ", messages)
+        new_conversation = save_to_firestore(st, messages)
+        if new_conversation is not None:
+            st.experimental_set_query_params(cid=new_conversation.id)
+            st.experimental_rerun()

render_conversation.py

@@ -27,5 +27,4 @@ def render_messages(st, messages):
 
 def render_conversation(st, conversation):
     messages = conversation["messages"]
-
     render_messages(st, messages)