The Problem: Info Overload from Endless Telegram Channels

You start with 10 channels. Then it becomes 50. Then you start marking things “read” just to feel less guilty. News gets duplicated, some posts get deleted, important stuff gets lost.

I’ve been through it myself. So I said: enough. We need something that collects information, understands what it’s about, and answers questions like:

• “What happened yesterday?”
• “Summarize this week’s news agenda.”
• “Who writes most often about tech/startups/economy?”
• “Which channels had spikes in negative sentiment this month?”
• “Show me all posts about the recent Apple keynote.”

What is RAG Search (and Why It’s Not Just Keyword Search)

RAG = Retrieval-Augmented Generation — an AI approach where the model doesn’t just “make things up” but pulls facts from a database and uses them in responses. It’s a combo of “search + generation.”

1. Take a user question — like “what was written about oil in May?”
2. Convert it into an embedding (a vector — basically a list of numbers)
3. Use Qdrant to find semantically similar texts (vector-to-vector match)
4. Pass results to GPT for answer generation

Result:

• No hallucinations — responses are grounded in real data;
• Fast and to the point;
• Understands meaning, not just word matches.

Code Example

    
async function getEmbedding(text) {
    const {
        data: [{ embedding }],
    } = await openai.embeddings.create({
        model: 'text-embedding-3-small',
        input: text,
        encoding_format: 'float',
    });
    return embedding;

/* This function sends text to OpenAI to get an embedding —
a numeric representation of meaning.
For example, “Oil prices rose” and “Crude became more expensive”
will have similar vectors. */
}
  

    
async function findSimilarPosts(query, top_k = 20) {
    const vector = await getEmbedding(query);
    const results = await qdrant.search('tg_posts', {
        vector,
        limit: top_k,
    });
    return results.map((r) => ({
        id: r.id,
        score: r.score,
        text: r.payload.text,
    }));

/* What this function does:

1. Converts the user's query into a vector.
2. Searches Qdrant for posts with semantically similar meaning.
3. Returns a list of posts ranked by similarity. */

}
  

RAG vs Keyword Search

Why This is Fast and Easy Now

Back then:

• Embeddings were expensive or just plain dumb;
• Vector databases like Qdrant/Pinecone were unstable;
• Telegram userbots were a pain to set up.

Now:

• OpenAI embeddings are cheap and high-quality;
• Qdrant/Pinecone are reliable and easy to integrate;
• Telegram.js or Telethon userbots take less than an hour to configure.

This lets you spin up a working MVP in a week — exactly how we built ours.

Our Pipeline

1. Telegram Userbot collects posts (even deleted ones leave traces);
2. Embedding + semantic search with Qdrant;
3. API serves responses, graphs, and tone analytics.

💬 FAQ

Tech Stack

• Node.js + Express
• Qdrant or Pinecone
• OpenAI embeddings
• Telegram.js / Telethon userbot
• React + Chart.js frontend

Want One of These?

We can build your MVP in 2–3 weeks. Connect your channels, tweak the visuals, no endless meetings or bureaucratic overhead.

Telegram: @madmaterials_bot

Email: team@madmaterials.com

Rustam

CEO & Team Lead