Cortex_LLM/Streamlit_Cortex_Evaluation_LLM.ipynb

{
 "cells": [
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Simple Streamlit App for Evaluation LLM\n",
    "\n",
    "Simple example showing one way to capture user feedback from using a LLM. Lots of ways to continue to improve and modify this.  \n",
    "\n",
    "A much better example can be found here: https://llm-eval-demo.streamlit.app/\n",
    "\n",
    "![Image](Streamlit_Eval.gif)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "I created a table `cortex_eval_table` with columns for ID, PROMPT, RESPONSE, MODEL, and TIMESTAMP\n",
    "\n",
    "```\n",
    "CREATE TABLE cortex_eval_table (\n",
    "    id INTEGER AUTOINCREMENT PRIMARY KEY,\n",
    "    prompt VARCHAR(2000),\n",
    "    model VARCHAR (255),\n",
    "    response VARCHAR(10000),\n",
    "    feedback VARCHAR(20),\n",
    "    timestamp TIMESTAMP_NTZ\n",
    ");\n",
    "```"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "# Import python packages\n",
    "import streamlit as st\n",
    "from datetime import datetime\n",
    "from streamlit import cache_resource, cache_data\n",
    "from snowflake.snowpark.session import Session\n",
    "from snowflake.snowpark.context import get_active_session\n",
    "import re\n",
    "import json\n",
    "\n",
    "session = get_active_session()\n",
    "\n",
    "@cache_data(show_spinner='CALLING LLM')\n",
    "def complete(model, prompt):\n",
    "    global session\n",
    "    prompt = re.sub(\"'\", \"''\", prompt)\n",
    "    sql_code = f\"SELECT SNOWFLAKE.CORTEX.COMPLETE('{model}', '{prompt}') as answer\"\n",
    "    df = session.sql(sql_code).to_pandas()\n",
    "    return df['ANSWER'][0], sql_code\n",
    "\n",
    "\n",
    "def update_feedback(session, record_id, feedback):\n",
    "    update_query = f\"\"\"\n",
    "    UPDATE cortex_eval_table\n",
    "    SET FEEDBACK = ?\n",
    "    WHERE id = ?\n",
    "    \"\"\"\n",
    "    session.sql(update_query, [feedback, record_id]).collect()\n",
    "\n",
    "st.header(\"Evaluation App with Cortex\")\n",
    "st.write(\"All prompts, responses, and feedback are saved.\")\n",
    "completion_input = st.text_area(\"Prompt\",height=200)\n",
    "completion_model = st.selectbox(\"Choose Model\", options = ['mixtral-8x7b','mistral-7b'])\n",
    "\n",
    "if 'record_id' not in st.session_state:\n",
    "    st.session_state.record_id = None\n",
    "if 'feedback' not in st.session_state:\n",
    "    st.session_state.feedback = None\n",
    "\n",
    "# Function to insert data and retrieve ID\n",
    "def insert_data_and_get_id(session, prompt, response, model):\n",
    "    formatted_timestamp = datetime.now().strftime('%Y-%m-%d %H:%M:%S')\n",
    "    #Update table\n",
    "    insert_query = f\"\"\"\n",
    "    INSERT INTO cortex_eval_table (PROMPT, RESPONSE, MODEL, TIMESTAMP, FEEDBACK)\n",
    "    VALUES (?, ?, ?, ?, ?)\n",
    "    \"\"\"\n",
    "    session.sql(insert_query, [prompt, response, model, formatted_timestamp, None]).collect()\n",
    "\n",
    "    ##Get ID number\n",
    "    result = session.sql(\"SELECT MAX(id) FROM cortex_eval_table\").collect()\n",
    "    new_id = result[0][0]  # Assuming MAX(id) gives us the newest ID safely\n",
    "    return new_id\n",
    "\n",
    "# Button to complete the task and save data\n",
    "if st.button(\"Complete\"):\n",
    "    if completion_input and completion_model:\n",
    "        result, sql_code = complete(completion_model, completion_input)\n",
    "        st.success(result)\n",
    "        record_id = insert_data_and_get_id(session, completion_input, result, completion_model)\n",
    "        st.session_state.record_id = record_id\n",
    "        st.success(\"Data saved successfully!\")\n",
    "        with st.expander(\"See SQL\"):\n",
    "            st.code(sql_code, language='SQL')\n",
    "\n",
    "# Feedback buttons\n",
    "col1, col2 = st.columns(2)\n",
    "if col1.button('👍', key='like'):\n",
    "    st.session_state.feedback = 'like'\n",
    "    update_feedback(session, st.session_state.record_id, st.session_state.feedback)\n",
    "    st.success(\"Thank you for your feedback!\", icon=\"✅\")\n",
    "\n",
    "if col2.button('👎', key='dislike'):\n",
    "    st.session_state.feedback = 'dislike'\n",
    "    update_feedback(session, st.session_state.record_id, st.session_state.feedback)\n",
    "    st.error(\"Thank you for your feedback!\", icon=\"✅\")\n",
    "\n",
    "# Display message on what feedback was registered\n",
    "if st.session_state.feedback is not None:\n",
    "    st.write(f\"Feedback {st.session_state.feedback} recorded at ID {st.session_state.record_id}\")     \n",
    "def fetch_data():\n",
    "    query = \"SELECT * FROM cortex_eval_table\"\n",
    "    return session.sql(query).to_pandas()\n",
    "\n",
    "# Display the results below\n",
    "st.write(\"## All Results\")\n",
    "results_df = fetch_data()\n",
    "st.dataframe(results_df)\n"
   ]
  }
 ],
 "metadata": {
  "kernelspec": {
   "display_name": "working38",
   "language": "python",
   "name": "python3"
  },
  "language_info": {
   "codemirror_mode": {
    "name": "ipython",
    "version": 3
   },
   "file_extension": ".py",
   "mimetype": "text/x-python",
   "name": "python",
   "nbconvert_exporter": "python",
   "pygments_lexer": "ipython3",
   "version": "3.8.18"
  }
 },
 "nbformat": 4,
 "nbformat_minor": 2
}