BardMind — Telegram Bot (Jacek Kaczmarski)

A Telegram bot that simulates the personality and works of Jacek Kaczmarski (1957–2004), the Polish bard and poet. Powered by GPT-4 with a RAG (Retrieval-Augmented Generation) pipeline built on OpenAI embeddings.

Architecture

Telegram Update
    └─▶ KaczmarskiBot (TelegramLongPollingBot)
            └─▶ MessageHandler
                    └─▶ CommandStorage
                            ├─▶ CommandExecutor      — predefined commands from commands.properties
                            ├─▶ ChatSessionManager   — tracks active chat sessions
                            └─▶ ChatHandler
                                    └─▶ KaczmarskiGPTHandler
                                            ├─▶ RagService       — context retrieval
                                            │       └─▶ VectorStore / EmbeddingService
                                            └─▶ OpenAiService    — GPT-4 / fallback GPT-3.5

Requirements

• Java 21+
• Maven 3.8+
• OpenAI API key
• Telegram bot token (from @BotFather)

Configuration

Copy the example config and fill in your credentials:

cp src/main/resources/config.properties.example src/main/resources/config.properties

openai.api.key=sk-...
telegram.bot.token=...
bot.username=YourBotName

The file is excluded from git (.gitignore).

Running

# 1. (Optional) Download Wikipedia articles about Kaczmarski into data/
mvn exec:java@import-wikipedia

# 2. Generate embeddings (once, or after adding new .txt files to src/main/resources/data/)
mvn exec:java@generate-embeddings

# 3. Start the bot
mvn exec:java@run-bot

By default the bot looks for embeddings.json in the working directory. Override with:

mvn exec:java@run-bot -Dbardmind.embeddings=/path/to/embeddings.json

Bot commands

Command	Description
/start	Start a conversation with Kaczmarski (activates GPT mode)
/end	End the active conversation
/help	List available commands
/info	About the bot
/nielubie	Easter egg

Once a session is started with /start, all messages are routed to GPT-4. Type daj tekst [title] to get the exact lyrics of a song.

How RAG Works

1. .txt files from src/main/resources/data/ are embedded using text-embedding-3-small.
2. Embeddings are saved to embeddings.json (external file, outside the JAR).
3. On each query: the question is embedded → cosine similarity search → top-K chunks → injected into the user message as context for GPT.

LLM Parameters (AppConfig.java)

Parameter	Value
Primary model	gpt-4
Fallback model	gpt-3.5-turbo
Temperature	0.3
Max tokens	300
Presence penalty	0.7
Frequency penalty	0.7
RAG top-K	2

Tests

mvn test

Tests require a valid OpenAI key in config.properties. Integration tests hit the real API and are slow (~10s).

Resource File Structure

src/main/resources/
├── config.properties          # NOT in repo — API keys
├── config.properties.example  # Template for new contributors
├── commands.properties        # Telegram commands and their actions
├── messages.properties        # UI messages (i18n-ready)
├── lyrics.properties          # Song lyrics (format: key=title|lyrics)
├── prompts/
│   └── system_prompt.txt      # System prompt for GPT
├── data/
│   ├── Mury.txt
│   ├── Rokosz.txt
│   └── Źródło.txt
└── embeddings.json            # Generated embeddings (not in repo)