katanemo/archgw
Archgw
Archgw ist ein Gateway-Proxy und eine Data-Plane, die für agentische Anwendungen konzipiert ist und eine zentrale Schicht für Routing, Sicherheit und Orchestrierung zwischen Anwendungslogik und mehreren Anbietern von Large Language Models bietet. Es fungiert als KI-Agent-Orchestrator, der die Ausführung von Agenten-Workflows automatisiert, um repetitive Aufgaben aus der Kern-Codebasis zu entfernen.
Das System verfügt über eine anbieterunabhängige Schnittstellenschicht, die unterschiedliche Modell-APIs in ein einheitliches Format standardisiert, sowie über eine transparente Proxy-Data-Plane zum Abfangen des Datenverkehrs. Es nutzt regelbasiertes Routing, um Anwendungslogik von spezifischen Anbietern zu entkoppeln, und verwendet Interceptor-basierte Guardrail-Ketten, um Ein- und Ausgaben auf Inhaltssicherheit und Jailbreak-Prävention zu filtern.
Die Plattform enthält ein Observability-System, das automatisch End-to-End-Traces, Metriken und Logs aus Modellinteraktionen erfasst, ohne dass eine manuelle Instrumentierung erforderlich ist. Zudem nutzt sie Middleware-gesteuerte Workflow-Orchestrierung, um den Fluss agentischer Aufgaben zu verwalten und Performance-Signale nachzuverfolgen.
Features
- Agentic Data Planes - Provides a centralized data plane for routing and safety between application logic and various AI model endpoints.
- Agentic Workflow Orchestration - Provides a system for executing complex agent processes by delegating tasks to autonomous agents.
- Model Request Routing - Provides mechanisms for directing API requests to different AI backends based on predefined routing rules.
- AI Safety Guardrails - Implements systems for detecting model jailbreaks, moderating content, and enforcing safety policies.
- Model Proxy Gateways - Implements a gateway proxy that provides a unified interface and routing for multiple LLM providers.
- Provider-Agnostic Model Interfaces - Provides an abstraction layer that standardizes disparate LLM provider APIs into a single consistent format.
- AI Agent Orchestrators - Coordinates model providers and task execution specifically to manage the lifecycle of AI agents.
- Signal-Based Routing Rules - Evaluates decision rules over signals to select the appropriate model provider and decouple application logic.
- Transparent Proxies - Implements a proxy that intercepts agent traffic for telemetry and policy enforcement without requiring client-side changes.
- Model Safety Filters - Ships validation layers for blocking or sanitizing model inputs and outputs based on safety policies.
- LLM Safety Enforcers - Enforces jailbreak, PII, and hallucination controls at the routing layer to block unsafe content.
- Input and Output Guardrails - Provides middleware to filter out-of-scope queries and validate generated content against safety rules.
- Request-Response Filter Chains - Ships a sequence of interceptors to validate and filter AI requests and responses for safety and correctness.
- Agent Workflow Orchestrators - Executes automated agent tasks by routing requests through programmable logic layers and state handlers.
- LLM Observability - Provides monitoring and tracing tools specifically tailored for LLM applications and agentic workflows.
- Agent Observability - Captures traces and evaluates the decision-making accuracy of autonomous agents without manual instrumentation.
- Agent Performance Monitoring - Tracks operational metrics and interaction signals from automated agents to monitor performance.
- Automatic Tracing Instrumentation - Implements automatic instrumentation for capturing end-to-end execution details and performance metrics for model calls.
- LLM Performance Monitoring - Automatically tracks performance metrics and execution traces for large language model operations.
- Agent Frameworks - High-performance proxy for agent routing and guardrail application.
- Agent Orchestration - Intelligent gateway for managing agent interactions.
