A minimal LLM chatbot for Vector built with:

• Vector SDK (Vector Bot client)
• dotenvy for environment config
• reqwest for OpenAI-compatible LLM calls
• Serde + std::fs for simple per-chat JSON memory stores

Core behavior:

• Per-chat memory: one JSON file per chat under data/, retaining the system prompt and up to N recent turns
• LLM calls use the OpenAI-compatible /chat/completions schema, so you can point it to OpenAI, llama.cpp servers, Groq, etc.
• Listens for gift-wrap (NIP-59) events, unwraps via Vector SDK, processes direct messages only, and replies privately

• Cargo.toml — crate metadata and dependencies
• src/main.rs — Vector bot bootstrap, subscription loop, unwrap and reply flow
• src/config.rs — loads configuration from environment variables
• src/llm.rs — OpenAI-compatible client
• src/memory.rs — simple per-chat JSON memory with trimming
• .env.example — template for configuration
• data/ — per-chat memory files (gitignored)

• Rust (stable) and Cargo installed
• Network access to your LLM endpoint (OpenAI, local llama.cpp, etc.)

Copy the example environment file and edit values:

cp .env.example .env

Environment variables (defaults shown):

• LLM_BASE_URL=https://api.openai.com/v1
  • Base URL for an OpenAI-compatible API. Examples:
    • OpenAI: https://api.openai.com/v1
    • llama.cpp server: http://localhost:8080/v1
    • Groq OpenAI compatible: https://api.groq.com/openai/v1
• LLM_API_KEY=sk-... (optional for local servers)
• LLM_MODEL=gpt-4o-mini (e.g., llama-3.1-8b-instruct)
• LLM_TEMPERATURE=0.2
• HISTORY_LIMIT=16 (number of user/assistant messages to keep per chat)
• DATA_DIR=data
• SYSTEM_PROMPT="You are a helpful assistant. Keep responses concise and factual."
• VECTOR_SECRET_KEY= (optional; nsec bech32 or hex). If unset, ephemeral keys are generated
• TYPING_INDICATOR=true (send kind 30078 typing indicator before replying)

cargo run

On start, the bot will:

1. Load environment config
2. Create keys (generate if VECTOR_SECRET_KEY is not set)
3. Create a VectorBot with default metadata and relays
4. Subscribe to gift-wrap events and process direct messages
5. For each message:
   • Load per-chat memory (JSON) and append the user message
   • Trim memory to HISTORY_LIMIT
   • Call your configured LLM (/chat/completions)
   • Persist the assistant reply and send it back privately

Run your llama.cpp server with OpenAI-compatible HTTP (ensure /v1/chat/completions is served):

# Example; your binary/flags may differ
./llama.cpp/server -m ./models/llama-3.1-8b-instruct.gguf --port 8080 --api

Use a .env like:

LLM_BASE_URL=http://localhost:8080/v1
LLM_MODEL=llama-3.1-8b-instruct
LLM_TEMPERATURE=0.2
HISTORY_LIMIT=12
DATA_DIR=data
TYPING_INDICATOR=true
# No API key required for local server

Then:

cargo run

LLM_BASE_URL=https://api.openai.com/v1
LLM_API_KEY=sk-...
LLM_MODEL=gpt-4o-mini
HISTORY_LIMIT=16
DATA_DIR=data
TYPING_INDICATOR=true

cargo run

For a chat ID (sender npub), the bot stores:

data/npub1abc...xyz.json

It contains:

• system prompt (string)
• history_limit (usize)
• messages: array of user/assistant messages

Oldest turns are trimmed when exceeding HISTORY_LIMIT. The system prompt is not counted against this limit.

• Unwrapping messages uses Vector SDK (no direct nostr-sdk usage required in your code).
• The bot filters only Kind::PrivateDirectMessage after unwrap and ignores other kinds.
• If TYPING_INDICATOR=true, it sends a brief typing indicator before LLM completion.

• Build errors about missing environment variables usually mean your .env isn’t loaded; confirm it exists and values are present.
• If HTTP 401/403 from LLM, confirm LLM_API_KEY and LLM_BASE_URL are correct.
• For local servers, confirm they implement /v1/chat/completions with the standard OpenAI payload/response shape.

MIT