lex-llm-gateway

Version: 0.2.15 | License: MIT | Ruby: >= 3.4

LLM inference gateway for LegionIO. Provides centralized metering over RabbitMQ, fleet RPC dispatch to GPU workers, and local disk spooling for offline resilience.

Add to your Gemfile:

gem 'lex-llm-gateway'

lex-llm-gateway wraps all LLM calls with automatic metering and fleet routing. It is designed for clusters with 100k+ edge nodes that cannot have direct database access.

Three node roles:

Full stack (transport + gateway + LLM + fleet)
  no transport  -> spool to disk, flush when reconnected
  no gateway    -> Legion::LLM direct (no metering)
  no fleet      -> local/cloud only
  no cloud      -> local LLM only
  no local      -> error

• Metering - build_event, publish_or_spool, flush_spool
• Inference - chat, embed, structured (all auto-metered)
• Fleet - dispatch to GPU workers with timeout and JWT auth
• FleetHandler - handle_fleet_request (validates JWT, calls local LLM)
• MeteringWriter - write_metering_record (DB insert consumed from RMQ)

require 'legion/extensions/llm/gateway/client'

client = Legion::Extensions::Llm::Gateway::Client.new
result = client.chat(model: 'claude-opus-4-6', messages: [{ role: 'user', content: 'Hello' }])
result[:success]  # => true
result[:response] # => "Hello! How can I help you?"

{
  "llm": {
    "routing": {
      "use_fleet": true,
      "fleet": {
        "timeout_seconds": 30,
        "require_auth": false
      }
    }
  }
}

• Ruby >= 3.4
• LegionIO framework
• legion-transport (AMQP metering + inference queues)
• legion-crypt (JWT signing for fleet auth, optional)
• legion-data (MeteringWriter and disk spool, optional)
• legion-llm (preferred inference, fleet, and metering owner for new work)

bundle install
bundle exec rspec     # 199 examples, 0 failures
bundle exec rubocop   # 0 offenses

MIT