Telegram AI Bot Development with LLM Context & Voice

A Telegram bot without context is useless. LLMs don't remember the dialog; each request is isolated. We solve this with FSM on Redis and automatic history compression. The result is a Telegram AI bot that remembers you, processes voice and images, and you control the budget. Backed by 5+ years of AI/ML expertise and over 30 delivered projects, our team ensures production-grade reliability.

Core Problems Addressed by a Telegram AI Bot

Dialog Context. LLMs don't remember history — each message is processed in isolation. We store the dialog in Redis via aiogram FSM. We limit history to the last 20 messages (configurable). For long sessions, we automatically compress context using summarization — compressing old messages into one summary.

Multimodal Input. Telegram supports text, voice, and photos. Voice is transcribed via Whisper (through Groq — latency <1 sec). Images are processed via vision API of Claude or GPT-4o. The user doesn't need to switch between tools — everything in one chat.

Rate Limiting and Budget. Each LLM request costs money (tokens). An attacker could spam and drain the account. We set a limit: no more than 10 requests per minute per user. When exceeded, the bot replies: "⏳ Too many requests. Please wait a minute." Additionally — a daily cap on tokens with admin notification.

Technical Stack and Implementation

Stack: aiogram 3.x (async framework), Redis (FSM and cache), Anthropic SDK (Claude), Groq SDK (Whisper), Anthropic Vision API. Configuration via environment variables, deployment in Docker.

Basic AI Bot with aiogram 3.x

import asyncio
from aiogram import Bot, Dispatcher, Router, types, F
from aiogram.filters import CommandStart, Command
from aiogram.fsm.context import FSMContext
from aiogram.fsm.storage.redis import RedisStorage
from anthropic import AsyncAnthropic

BOT_TOKEN = "BOT_TOKEN"
ANTHROPIC_API_KEY = "ANTHROPIC_API_KEY"

bot = Bot(token=BOT_TOKEN)
storage = RedisStorage.from_url("redis://localhost:6379")
dp = Dispatcher(storage=storage)
router = Router()
dp.include_router(router)
thropic_client = AsyncAnthropic(api_key=ANTHROPIC_API_KEY)

# Dialog history stored in FSM context
@router.message(CommandStart())
async def start(message: types.Message, state: FSMContext):
    await state.update_data(history=[])
    await message.answer(
        "Hi! I'm an AI assistant based on Claude. Ask me anything.",
        reply_markup=get_main_keyboard()
    )

def get_main_keyboard():
    from aiogram.utils.keyboard import ReplyKeyboardBuilder
    builder = ReplyKeyboardBuilder()
    builder.button(text="🗑 Clear history")
    builder.button(text="ℹ️ About bot")
    builder.adjust(2)
    return builder.as_markup(resize_keyboard=True)

@router.message(F.text == "🗑 Clear history")
async def clear_history(message: types.Message, state: FSMContext):
    await state.update_data(history=[])
    await message.answer("Dialog history cleared.")

@router.message(F.text & ~F.text.startswith("/"))
async def handle_text(message: types.Message, state: FSMContext):
    data = await state.get_data()
    history = data.get("history", [])

    # Add user message to history
    history.append({"role": "user", "content": message.text})

    # Show typing
    await bot.send_chat_action(message.chat.id, "typing")

    # Streaming response
    full_response = ""
    sent_message = await message.answer("...")

    async with anthropic_client.messages.stream(
        model="claude-haiku-4-5",
        max_tokens=2048,
        system="You are a helpful assistant. Answer concisely and to the point.",
        messages=history,
    ) as stream:
        async for text in stream.text_stream:
            full_response += text
            # Update message every 50 characters
            if len(full_response) % 50 == 0:
                await sent_message.edit_text(full_response)

    await sent_message.edit_text(full_response)

    # Save response to history
    history.append({"role": "assistant", "content": full_response})
    # Keep last 20 messages
    await state.update_data(history=history[-20:])

Voice Message Handling

import io
from groq import AsyncGroq

groq_client = AsyncGroq(api_key="GROQ_API_KEY")

@router.message(F.voice)
async def handle_voice(message: types.Message, state: FSMContext):
    await bot.send_chat_action(message.chat.id, "typing")

    # Download voice message
    voice = message.voice
    file = await bot.get_file(voice.file_id)
    file_bytes = await bot.download_file(file.file_path)

    # Transcribe with Whisper on Groq (fast)
    transcription = await groq_client.audio.transcriptions.create(
        file=("voice.ogg", file_bytes.read()),
        model="whisper-large-v3",
        language="en",
    )

    text = transcription.text
    await message.answer(f"📝 Recognized: {text}")

    # Process as text message
    await handle_text_with_content(message, state, text)

Image Handling

import base64

@router.message(F.photo)
async def handle_photo(message: types.Message, state: FSMContext):
    await bot.send_chat_action(message.chat.id, "typing")

    # Get the best quality photo
    photo = message.photo[-1]
    file = await bot.get_file(photo.file_id)
    file_bytes = await bot.download_file(file.file_path)

    image_data = base64.standard_b64encode(file_bytes.read()).decode()
    caption = message.caption or "What is in this image?"

    response = await anthropic_client.messages.create(
        model="claude-haiku-4-5",
        max_tokens=1024,
        messages=[{
            "role": "user",
            "content": [
                {"type": "image", "source": {"type": "base64", "media_type": "image/jpeg", "data": image_data}},
                {"type": "text", "text": caption},
            ]
        }]
    )

    await message.answer(response.content[0].text)

Rate Limiting and Billing Protection

from aiogram.filters import BaseFilter
from collections import defaultdict
import time

class RateLimitFilter(BaseFilter):
    def __init__(self, rate_limit: int = 10, period: int = 60):
        self.rate_limit = rate_limit
        self.period = period
        self.user_requests = defaultdict(list)

    async def __call__(self, message: types.Message) -> bool:
        user_id = message.from_user.id
        now = time.time()

        # Remove old requests
        self.user_requests[user_id] = [
            t for t in self.user_requests[user_id]
            if now - t < self.period
        ]

        if len(self.user_requests[user_id]) >= self.rate_limit:
            await message.answer("⏳ Too many requests. Please wait a minute.")
            return False

        self.user_requests[user_id].append(now)
        return True

# Apply the filter
@router.message(RateLimitFilter(rate_limit=10), F.text)
async def handle_with_limit(message: types.Message, state: FSMContext):
    await handle_text(message, state)

Ensuring Dialog Context in a Telegram Bot

For persistent context, we use Redis Storage. FSM state is saved between requests. The dialog history is an array of messages with user and assistant roles. We limit it to the last 20 messages to avoid exceeding the model's context window (Claude Haiku has 200k tokens). If the dialog is longer, we automatically compress old messages into one summary via LLM.

Why Redis Instead of PostgreSQL?

Redis gives latency p99 < 10 ms, PostgreSQL — ~50 ms. For FSM, where each request rewrites state, speed is critical. Also, Redis automatically removes outdated data (TTL), simplifying maintenance.

Development Process

1. Requirements analysis (1 day): Technical specification, LLM selection, architecture design.
2. Bot prototype (2–3 days): MVP with text-based dialog using FSM and Redis.
3. Voice/image integration (2–3 days): Add multimodal input via Whisper and vision API.
4. Rate limiting and admin panel (3–5 days): Implement abuse protection, budget caps, Prometheus monitoring.
5. Deployment and testing (2–3 days): Release on VPS with Docker, load testing.

LLM Model Comparison for Telegram Bot

Claude Haiku is 2.5x faster than GPT-4o with comparable quality for typical tasks, making it optimal for chatbots.

Practical Case: Customer Support

From our practice: we developed a Telegram bot on Claude Haiku for a SaaS product. Result: 73% of inquiries handled without human support. Response time: from 24 hours to instant. The client saved approximately $800 per month on support costs. We configured rate limiting and a daily token cap — daily spend did not exceed $1.5, confirming the approach's effectiveness.

Common Mistakes in AI Bot Development

• Storing history in process memory. Context is lost on bot restart. Use Redis.
• Ignoring rate limiting. A single user can spam and drain the budget. Always set limits.
• No error handling for LLM calls. The model may crash or return an invalid response. Wrap calls in try-except.
• Hard binding to one model. If the model becomes unavailable, the bot stops working. Plan a fallback to another LLM.

What's Included

• Technical specification and stack selection.
• Bot development (dialog, voice, images).
• Rate limiting and admin configuration.
• Redis integration and Docker deployment.
• Documentation, access, and team training.
• One month of support after launch.

Estimated Timeline

• Basic bot with dialog: 2–3 days
• Voice + images: 2–3 days
• Rate limiting + admin panel: 1 week
• Release + monitoring: +2–3 days

Contact us for a free consultation and accurate cost estimate. Order turnkey development with quality guarantee.

LLM Development: Fine-Tuning, RAG, Agents, and Production Deployment

Using GPT‑4 or Claude 3.5 Sonnet through a public API is not a solution — it's just a tool. When the requirement is to "make it like ChatGPT, but on our data," there is a real engineering challenge behind it: from prompt engineering to training a 70B model on your own infrastructure. End-to-end LLM solution development is a complex stack, and we have been doing it for over 5 years. During this time, we have completed over 20 projects in generative AI: from RAG systems for legal departments to custom support agents. Where exactly your task falls depends on data, latency requirements, budget, and how critical confidentiality is.

A typical situation: the client has already tried ChatGPT, but results are unstable — sometimes accurate, sometimes hallucinating. Or they need integration into a corporate portal while complying with security policies. Let's break down each layer of the stack in detail — from RAG to production deployment.

Why Do RAG Systems Break and How to Fix It?

RAG (Retrieval-Augmented Generation) looks simple: find relevant documents, put them in context, get an answer. In practice, it fails in several places.

Chunking without overlap. Classic mistake: chunk_size=512, overlap=0. If the answer lies across two chunks, retrieval won't find either with sufficient confidence. Solution: overlap 15–25% of chunk_size, or better yet, sentence-aware splitting with spaCy or NLTK instead of naive character splitting.

Poor embedder. text-embedding-ada-002 is good for general use, but on legal or medical texts, specialized models like E5-large-v2, BGE-M3, or fine-tuned sentence-transformers on domain data outperform it. Recall@5 differences can be 15–25%.

No re-ranking. Vector search optimizes for speed, not relevance. A cross-encoder re-ranker (ms-marco-MiniLM-L-6-v2, bge-reranker-large) after initial retrieval improves top-3 accuracy with acceptable latency (+50–150ms). This is often more impactful than improving the embedding model.

Hybrid search. Dense vectors alone work poorly on exact queries: names, SKUs, codes. BM25 (sparse) finds exact matches but misses semantics. Hybrid via RRF (Reciprocal Rank Fusion) is the optimal compromise. Qdrant, Weaviate, and pgvector 0.7+ support hybrid search natively.

When to Fine-Tune Instead of Prompt Engineering?

Prompt engineering solves ~70% of LLM adaptation tasks for a domain. The remaining 30% require fine-tuning. Three indicators: the model ignores a specific output format even with detailed prompting; the task requires deep knowledge of specialized vocabulary (medicine, law); you need to significantly reduce token costs by replacing a large model with a smaller specialized one.

LoRA and QLoRA are the standard for SFT. LoRA adds trainable low-rank matrices to attention layers. A typical configuration for Llama-3 8B: r=64, lora_alpha=128, target_modules=["q_proj","v_proj","k_proj","o_proj"] yields ~0.8% trainable parameters, training on one A100 40GB. QLoRA adds 4-bit quantization (NF4) and allows fine-tuning 70B models on two A100 40GB, though speed drops by half compared to bf16.

DPO instead of RLHF. Direct Preference Optimization requires only (chosen, rejected) pairs, not scalar reward signals. DPOTrainer from the trl library (Hugging Face) implements it in a few dozen lines.

Common mistake. A dataset of 500 examples, 5 epochs, validation loss 0.8 — seems fine. But on test, the model degrades on general instructions. Cause: catastrophic forgetting. Solution: add 10–20% general instruction-following examples (Alpaca, FLAN) to the training set to preserve original capabilities.

How to Choose a Base Model: 8B or 70B?

For most tasks, fine-tuning an 8B model is sufficient. 70B is needed when deep reasoning is required or the 8B baseline does not reach the required quality even after fine-tuning. Inference cost for Llama-3 8B via vLLM on A100 is efficient; the exact cost depends on volume.

What Does PagedAttention Bring to Production?

vLLM is the first choice for serving open-source models. PagedAttention is the key technical innovation: KV-cache is managed like virtual memory in an OS, without fragmentation. This yields 2–4x higher throughput compared to naive HuggingFace Transformers inference. The vLLM documentation confirms that continuous batching and PagedAttention are the standard for high-load LLM services.

Typical numbers on A100 80GB for Llama-3 8B (bf16): 400–600 req/s, P50 latency 200–400ms, P99 latency 600–900ms at concurrency 64. For 70B on two A100 with tensor parallelism: 80–120 req/s, P99 latency 1.5–2.5s. AWQ or GPTQ quantization reduces memory consumption by 2x with quality loss within 1–3%.

Multi-Agent Systems

Agents are LLMs with access to tools: search, code execution, API calls, database interaction. Common patterns:

• ReAct (Reason + Act): the model reasons → chooses a tool → observes the result → reasons again. LangChain and LlamaIndex implement it out of the box.
• Multi-agent orchestration: multiple specialized agents with a coordinator on top. Example: coordinator → researcher (search + summarization) → coder (code generation and execution) → critic (verification). Tools: AutoGen (Microsoft), CrewAI, custom implementation on LangGraph.

In production, agent systems are non-deterministic. Essential: guardrails, step limits, logging of each step, human-in-the-loop for critical actions.

How We Work: Stages, Timeline, Deliverables

What Is Included

We deliver:

• Technical documentation (model card, configs, deployment instructions)
• Access to infrastructure (code repository, trained weights)
• 1 month of post-deployment support (consultations, bug fixes)
• Customer team training (2–3 sessions on system operation)

Timeline: basic RAG prototype — 1–2 weeks. Fine-tuning with customer data — 3–6 weeks (including data preparation). Production system with monitoring and retraining — 2–4 months. Cost is calculated individually based on data volume, model complexity, and infrastructure requirements.

We guarantee the quality of the final model with performance benchmarks and ongoing monitoring. Our engineers have hands‑on experience with dozens of production LLM systems.

Want to evaluate your project? Leave a request — we will prepare a preliminary summary within 1–2 business days. Or get a consultation on choosing the approach: RAG, fine-tuning, or hybrid — we will tell you what works best for you. Contact us to discuss your LLM development needs. Schedule a free consultation today.