The guts of l3mcore
Find out what makes this middleware so special.
Smart Routing
At the heart of l3mcore beats an engine of Local Machine Learning based on dense HuggingFace models. Instead of just searching for keywords, l3mcore converts your message into a mathematical vector and compares it with the semantic space of each expert configured in milliseconds. If the model doubts, an advanced algorithm Fuzzy Matching comes in as a backup.
Integrated Multi-Backend and Extreme Efficiency
Why limit yourself to just one technology? l3mcore is agnostic to the underlying engine. Can wake up a model ONNX local loading it into RAM with an LRU cache system for instant commands, fire a REST request to Ollama, or send the traffic to external APIs. Super optimized systems: in stress tests with 15 experts available in the system, the core consumes only 1.5GB RAM.
Security by Design (Open Source)
Being from open source and auditable, we guarantee total transparency. In exposed environments, l3mcore acts as a firewall that intercepts and blocks security vulnerabilities. Path Traversal y SSRF, preventing the injection of arbitrary routes. Additionally, it filters the maximum payload size and obfuscates any sensitive information in the logs to ensure regulatory compliance.
100% Compatible API
You don't have to reschedule your clients. l3mcore exposes a Flask server whose main endpoint mimics the industry standard. If you use a Frontend, AnythingLLM or the official Python library itself, you just need to change the connection URL to http://localhost:11435. l3mcore will act as a transparent translator.
Plugin System
Extend l3mcore capabilities to your needs. The new plugin system allows you to add extra functionalities without modifying the core code. Discover, download and create custom modules in the Plugin Directory.