A primitive NL to SQL chat feature

README.md

@@ -15,6 +15,7 @@ Your texting data never leaves your computer. We are proudly open-source for thi
 - 🔍 filter by a word, friend, or time range
 - 💯 sentiment analysis
 - 🎁 "Your Year in Text" experience a.k.a iMessage Wrapped
+- 💬 Chat with an LLM to query your data
 
 ## Download Left on Read for Mac
 

app/README.md

@@ -116,9 +116,25 @@ You most likely have an old electron process running. If this is warning, you ca
 
 If you are getting a "Cannot find module" error, you likely forgot to install the packages. Be sure to run `yarn` in the `app/` directory.
 
-
 ## Support
 
 <a href="https://www.buymeacoffee.com/leftonread" target="_blank"><img src="https://cdn.buymeacoffee.com/buttons/v2/default-yellow.png" alt="Buy Me A Coffee" style="height: 60px !important;width: 217px !important;" ></a>
 
 <a href="https://www.producthunt.com/posts/left-on-read?utm_source=badge-featured&utm_medium=badge&utm_souce=badge-left&#0045;on&#0045;read" target="_blank"><img src="https://api.producthunt.com/widgets/embed-image/v1/featured.svg?post_id=358899&theme=light" alt="Left&#0032;on&#0032;Read - iMessages&#0032;supercharged | Product Hunt" style="width: 250px; height: 54px;" width="250" height="54" /></a>
+
+## LLM Chat (first go)
+
+How it works:
+
+1. User inputs a natural language query (How many texts did I send to Mom?)
+2. This is passed to LLM to turn it into a SQL Query to run on core_main_table TODO: - give the LLM more of the DB schema
+3. Execute the generated query and synthesize query result + original message/query with the LLM to output readable response to the user
+
+Most of the backend logic happens in app/src/analysis/queries/RAGEngine.ts
+
+Notes:
+
+- Currently does not account for messages in group chat vs. not in a group chat, so numbers are misleading.
+- Must use the exact spelling of your contacts' name as it appears in the address book.
+- Requires an OpenAI API Key and internet access
+- Errors (including but not limited to internet connectivity, malformed SQL queries generated by LLM, etc.) are not all handled gracefully (some are OK, like invalid API key)

app/package.json

@@ -142,6 +142,7 @@
     "node-machine-id": "^1.1.12",
     "node-schedule": "^2.1.0",
     "nodemailer": "^6.7.7",
+    "openai": "^4.28.4",
     "path-browserify": "^1.0.1",
     "react": "^18.1.0",
     "react-chartjs-2": "^4.3.1",

app/src/analysis/queries/RagEngine.ts

@@ -0,0 +1,116 @@
+import OpenAI from 'openai';
+import * as sqlite3 from 'sqlite3';
+
+import * as sqlite3Wrapper from '../../utils/sqliteWrapper';
+
+// A test function for me to understand the altered DB schema better :)
+export async function printDBTableNames(
+  db: sqlite3.Database
+): Promise<string[]> {
+  const q = `
+  SELECT 
+    name 
+  FROM 
+    sqlite_master 
+  WHERE 
+    type='table' 
+  ORDER BY 
+    name
+  `;
+  return sqlite3Wrapper.allP(db, q);
+}
+
+// Put together a really hacky RAG pipeline...
+export async function queryRAGEngine(
+  db: sqlite3.Database,
+  message: string,
+  key: string
+): Promise<string> {
+  const openai = new OpenAI({
+    apiKey: key,
+  });
+
+  let q: string | null = null;
+
+  let prompt = `
+  Write a query for a table called core_main_table with this schema: 
+  contact_name, 
+  text (which has the message's text), 
+  date (a unix timestamp number in nanoseconds when the message was sent),
+  is_from_me (a boolean indicating if I was the sender of the message) 
+
+  to answer the following query: ${message}
+  Please respond with only the raw unformatted SQL and no other text. If this is not possible, or it's hard to get a concrete result based on the schema, return 'Not Possible'
+  `;
+
+  try {
+    const response = await openai.chat.completions.create({
+      model: 'gpt-3.5-turbo', // Change the model as per your requirement
+      messages: [{ role: 'system', content: prompt }],
+      temperature: 0.7,
+      max_tokens: 150,
+    });
+    q = response.choices[0].message.content;
+    console.log(response.choices[0]);
+  } catch (error) {
+    console.error(error);
+    return new Promise<string>((resolve, reject) => {
+      resolve('An error occurred. Check your API key and try a new message.');
+    });
+  }
+
+  const query = `
+  SELECT COUNT(*) AS message_count
+  FROM core_main_table
+  WHERE LOWER(contact_name) = LOWER('${message}');
+  `;
+
+  const queryResult = await sqlite3Wrapper.allP(db, q ?? query);
+
+  function isObject(value: any): value is Record<string, any> {
+    return (
+      typeof value === 'object' &&
+      value !== null &&
+      !Array.isArray(value) &&
+      !(value instanceof Date)
+    );
+  }
+  if (!isObject(queryResult[0])) {
+    console.log(queryResult[0]);
+  }
+  const resultString = JSON.stringify(queryResult[0]);
+  // Sanity check so you don't use don't accidentally use too many tokens...
+  if (resultString.length > 10000) {
+    return new Promise<string>((resolve, reject) => {
+      resolve('An error occurred. Try a new message.');
+    });
+  }
+
+  prompt = `
+  Given this message from a user: ${message}, 
+  this corresponding generated query over a database: ${query}, 
+  and this result of the query ${resultString}:
+  interpret the result of the query in plain english as a response to the initial message.
+  `;
+
+  let result = '';
+  try {
+    const response = await openai.chat.completions.create({
+      model: 'gpt-3.5-turbo', // Change the model as per your requirement
+      messages: [{ role: 'system', content: prompt }],
+      temperature: 0.7,
+      max_tokens: 150,
+    });
+    result = response.choices[0].message.content ?? 'An error occurred';
+    console.log(response.choices[0]);
+  } catch (error) {
+    console.error(error);
+    return new Promise<string>((resolve, reject) => {
+      resolve('An error occurred. Check your API key and try a new message.');
+    });
+  }
+
+  return new Promise<string>((resolve, reject) => {
+    resolve(result); // Resolve the promise with a string value
+  });
+}

app/src/components/Dashboard/ChatInterface.tsx

@@ -0,0 +1,114 @@
+import { Box, Button, Flex, Input, Text } from '@chakra-ui/react';
+import { ipcRenderer } from 'electron';
+import React, { useEffect, useRef, useState } from 'react';
+
+interface Message {
+  text: string;
+  sender: 'user' | 'bot';
+}
+
+const initialBotMessage: Message = {
+  text: 'Hi there :) You can ask me questions here about your iMessages! For example, try "Who is my best friend?"',
+  sender: 'bot',
+};
+
+interface ChatInterfaceProps {
+  openAIKey: string;
+}
+
+export function ChatInterface(props: ChatInterfaceProps) {
+  const { openAIKey } = props;
+
+  const [messages, setMessages] = useState<Message[]>([initialBotMessage]);
+  const [newMessage, setNewMessage] = useState<string>('');
+
+  const [awaitingResponse, setAwaitingResponse] = useState<boolean>(false);
+
+  const messagesContainerRef = useRef<HTMLDivElement>(null);
+
+  const handleMessageChange = (e: React.ChangeEvent<HTMLInputElement>) => {
+    setNewMessage(e.target.value);
+  };
+
+  const handleSendMessage = async () => {
+    if (newMessage.trim()) {
+      setMessages([...messages, { text: newMessage, sender: 'user' }]);
+      setNewMessage('');
+      setAwaitingResponse(true);
+
+      const llmResponse: string = await ipcRenderer.invoke(
+        'rag-engine',
+        newMessage,
+        openAIKey
+      );
+
+      setMessages((prevMessages) => [
+        ...prevMessages,
+        { text: llmResponse, sender: 'bot' },
+      ]);
+      setAwaitingResponse(false);
+    }
+  };
+
+  useEffect(() => {
+    if (messagesContainerRef.current) {
+      messagesContainerRef.current.scrollTop =
+        messagesContainerRef.current.scrollHeight;
+    }
+  }, [messages]);
+
+  return (
+    <Box width="90%" mx="auto" mt={8}>
+      <Box
+        ref={messagesContainerRef}
+        bg="gray.100"
+        p={4}
+        borderRadius="md"
+        height="inherit"
+        maxH="70vh"
+        minH="70vh"
+        overflowY="scroll"
+        display="flex"
+        flexDirection="column"
+      >
+        {messages.map((message, index) => (
+          <Flex
+            key={index}
+            mb={2}
+            alignItems="flex-end"
+            justifyContent={
+              message.sender === 'user' ? 'flex-end' : 'flex-start'
+            }
+          >
+            <Box
+              bg={message.sender === 'user' ? 'blue.500' : 'gray.300'}
+              color={message.sender === 'user' ? 'white' : 'black'}
+              p={2}
+              borderRadius="md"
+              maxW="80%"
+            >
+              <Text>{message.text}</Text>
+            </Box>
+          </Flex>
+        ))}
+      </Box>
+      <Flex mt={4}>
+        <Input
+          value={newMessage}
+          onChange={handleMessageChange}
+          placeholder="Type your message..."
+          mr={2}
+        />
+        <Button
+          colorScheme="purple"
+          onClick={handleSendMessage}
+          isLoading={awaitingResponse}
+        >
+          Send
+        </Button>
+      </Flex>
+    </Box>
+  );
+}
+
+export default ChatInterface;