Semantic search in Telegram bot for communities

Telegram Bot with Semantic Search for Streamlining Business Communities

In a community of 120 entrepreneurs, manually finding experts took hours, event sign-ups required back-and-forth DMs, and activity rankings were tracked in spreadsheets. This AI-powered Telegram bot eliminates those bottlenecks: vector search through member profiles finds specialists by query intent in seconds, rankings update automatically, and event registration is one-click. The system leverages FAISS for vector storage and GPT for text structuring, delivering spot-on results without relying on keywords.

Growth Pains: Three Key Bottlenecks

As the community scaled, these systemic issues emerged:

• Expert Search: Impossible to remember 100+ specializations. Chat queries like "anyone in logistics?" got buried, and the right person never saw them.
• Event Management: 5–7 hours weekly on lists, reminders, and tweaks. It just didn't scale.
• Activity Tracking: Manual spreadsheets only captured obvious "thanks," missing real contributions.

These are common hurdles for chat-based communities. The fix? Shift from keyword search to vector-based profile representations.

AI Search Architecture: From Text to Vectors

Profile Completion and Structuring

Each member submits free-form text: experience, skills, case studies. A GPT model (powered by OpenAI) parses it into three weighted layers:

• Core Summary (10–15 words, 55% weight): Key traits.
• Detailed Skills (30% weight): Full skill set.
• Context (15% weight): Hobbies, interests.

This handles varied writing styles. For instance, "visa help" matches profiles mentioning "business relocation."

Vectorization and Storage

Each layer becomes a vector (~3000 dimensions) via an embedding model. Vectors are stored in FAISS, a library for lightning-fast approximate nearest neighbor search.

The workflow:

• User query (e.g., "Essence: visas to Italy") gets vectorized.
• FAISS retrieves 10–15 closest profiles.
• An "auditor" model checks relevance, ranks by activity score.

Outcome: Top-3 matches with high semantic similarity, even without exact matches.

Automated Ranking with Telethon and PostgreSQL

The bot monitors chat via Telethon:

• "Thanks" = +1 point.
• Organizing a mastermind: +100.
• Event attendance: +30.

Data aggregates in PostgreSQL. Rankings boost:

• Search priority.
• Access to exclusive events.

This builds transparent reputation: active members rise to the top.

# Sample monitoring pseudocode (Python + Telethon)
from telethon import TelegramClient

async def monitor_thanks(client, chat_id):
    async for message in client.iter_messages(chat_id):
        if 'thanks' in message.text.lower():
            add_karma(message.reply_to_msg_id, 1)

Event Calendar: From Manual Lists to Self-Service

Bot-integrated calendar: announcements → "Sign up" button → auto-add to PostgreSQL. No messaging needed; reminders are bot-generated. Scales effortlessly to hundreds of members.

Key Takeaways

• FAISS + GPT semantic search hits 80–90% accuracy without keywords.
• Multi-layer vector profiles (with weights) handle unstructured text.
• Telethon-powered rating automation cuts manual work by 100%.
• Calendar integration simplifies events 10x.
• Works for any Telegram group: IT teams, business clubs, you name it.

— Editorial Team