Use AI to build your house!

When a new housing society emerges, residents inevitably create chat groups to connect and share information using various chat apps like WhatsApp and Telegram. In India, Telegram seems to be the favorite as it provides generous group limits, admin tools, among other features. These virtual communities become treasure troves of invaluable insights. But whatever app you use, there is always a problem of finding the right information at right time. Sure, the apps have a "Search" button, but they are pretty much limited to keyword search and are useless when you have to search through thousands of messages.

I found myself in this situation when it was my turn to start on an interior design project for my home. Despite being part of a vibrant Telegram group, where countless residents had shared their experiences with various interior designers and companies, I struggled to unearth the pearls of wisdom buried within the chat's depths. I remembered that I could take advantage of AI, particularly LLMs to sift through this data.

Telegram has a nifty feature where you can request for a data dump of a group that you are part of. The data is generally available 24 hours after request. Since my group had more than 50,000 messages, the data was shared in multiple HTML files containing the user chats. 

Telegram Group Data Dump

I needed to extract this into a simple text document that I can feed to an LLM.

I wrote a simple C# program to do this:

string? location;

do

{

    Console.WriteLine("Enter the location where Telegram chats are stored...");

    location = Console.ReadLine();

    if (string.IsNullOrWhiteSpace(location))

    {

        Console.WriteLine("Invalid location");

    }

} while (string.IsNullOrWhiteSpace(location));

//find all html files in the location

string[] htmlFiles = Directory.GetFiles(location, "*.html");

Console.WriteLine("Found " + htmlFiles.Length + " files");

foreach (string htmlFile in htmlFiles)

{

    Console.WriteLine($"Processing {htmlFile}");

    //load html file

    string html = File.ReadAllText(htmlFile);

    //extract text from div

    List<string> allChats = ExtractTextFromDiv(html);

    //write text to file

    File.AppendAllLines("chatDump.txt", allChats);

    Console.WriteLine($"Wrote {allChats.Count} chats.");

}

static List<string> ExtractTextFromDiv(string html)

{

    List<string> allChats = [];

    // Load the HTML document

    var doc = new HtmlDocument();

    doc.LoadHtml(html);

    // Find the div with class "text"

    var textDiv = doc.DocumentNode.SelectNodes("//div[@class='text']");

    if (textDiv == null || textDiv.Count == 0)

    {

        Console.WriteLine("No text div found");

        return allChats;

    }

    foreach (var node in textDiv)

    {

        if (!string.IsNullOrWhiteSpace(node.InnerText) && node.InnerText.Contains("interior"))

            allChats.Add(node.InnerText);

    }

    return allChats;

}

Finally, I could narrow it down to ~ 1300 messages. Next, I chunked these messages further into 200 messages each to avoid the context token limit of Anthropic's Claude Sonnet LLM and used Amazon Bedrock to extract the information.

static async Task<string> InvokeClaudeAsync(string messages)

{

    string claudeModelId = "anthropic.claude-3-sonnet-20240229-v1:0";

    string prompt = $"""

        You are provided with chat messages below. Find out the name and phone numbers

        of interior companies where users had a good experience

        and the work was completed with good quality and on time.

        Messages start below:

        {messages}

        """;

    // Claude requires you to enclose the prompt as follows:

    string enclosedPrompt = "Human: " + prompt + "\n\nAssistant:";

    AmazonBedrockRuntimeClient client = new(RegionEndpoint.USEast1);

    string payload = new JsonObject()

            {

                { "prompt", enclosedPrompt },

                { "max_tokens_to_sample", 200000 },

                { "temperature", 0.5 },

                { "stop_sequences", new JsonArray("\n\nHuman:") }

            }.ToJsonString();

    string generatedText = "";

    try

    {

        InvokeModelResponse response = await client.InvokeModelAsync(new InvokeModelRequest()

        {

            ModelId = claudeModelId,

            Body = AWSSDKUtils.GenerateMemoryStreamFromString(payload),

            ContentType = "application/json",

            Accept = "application/json"

        });

        if (response.HttpStatusCode == System.Net.HttpStatusCode.OK)

        {

            return JsonNode.ParseAsync(response.Body).Result?["completion"]?.GetValue<string>() ?? "";

        }

        else

        {

            Console.WriteLine("InvokeModelAsync failed with status code " + response.HttpStatusCode);

        }

    }

    catch (AmazonBedrockRuntimeException e)

    {

        Console.WriteLine(e.Message);

    }

    return generatedText;

}

Now I got a neat summary of information for each batch:

Based on the chat messages, here are the names and phone numbers of interior companies or persons who received positive feedback for good quality work delivered on time:

1. Redacted - One user mentioned finalizing interiors with them.

2. Redacted - A user had a pleasant experience with them for interiors at Redacted.

3. Redacted - A user's friend got good interiors done by them at Redacted.

4. Redacted - Recommended by a user for their great work on a villa.

5. Redacted - Quoted around 17L for a 2.5 BHK at Redacted, and one user liked their work.

6. Redacted - Did the model flat interiors at Redacted. A user mentioned they are open to discounts for bulk orders.

Some other companies mentioned without specific feedback were Redacted, Redacted and Redacted. A few users cautioned against Redacted based on poor experiences at Redacted.

The results were nothing short of remarkable. I received a meticulously curated summary, complete with names, phone numbers, and feedback on interior companies that had garnered praise for their quality work and timely delivery. Unexpectedly, the LLM even went the extra mile, cautioning against companies with poor track records, ensuring I had a well-rounded perspective. Now I had 7 set of responses which I fed again to the LLM to generate a neat list of good and bad companies!

Was using an LLM an overkill for this task? Perhaps. Was it fun? You bet!