SAGAI-MID: Using Large Language Models to Solve Dynamic Interoperability Challenges in Distributed Systems

The MIT team proposes the SAGAI-MID middleware, which uses large language models to dynamically detect and resolve API schema mismatches at runtime, enabling automatic adaptation of REST, GraphQL, and IoT devices with an accuracy rate of 90%. This addresses the pain point where traditional static adaptation solutions cannot handle new runtime scenarios.

In modern distributed systems, schema mismatches between heterogeneous services (different REST versions, GraphQL endpoints, IoT proprietary formats) hinder data flow. Traditional solutions require manual adapter writing and cannot handle new runtime combination scenarios; static adapters struggle to meet the dynamic access demands brought by the explosive growth of IoT devices; existing architectural tactics mostly stay at the design guidance level, lacking runtime automatic execution mechanisms.

SAGAI-MID is built on FastAPI and adopts a five-layer pipeline architecture:

1. Hybrid Detection Layer: Combines structural comparison and large language model semantic analysis to identify fields that are superficially different but semantically compatible (e.g., "user_id" and "customer_identifier");
2. Dual-Strategy Parsing Layer: Direct conversion (real-time data conversion, flexible) and code generation (generates reusable adapters, more efficient, with pass@1 accuracy of 0.83 which is better than direct conversion's 0.77);
3. Three-Layer Security Protection: Verification mechanism (schema check of converted data), integrated voting (multi-model consensus to improve reliability), rule fallback (predefined policies are enabled when model output is uncertain).

Tested across 10 scenarios (REST version migration, IoT-to-analytics platform bridging, GraphQL conversion, etc.) using 6 different large language models:

• The best configuration achieved an accuracy rate of 90%;
• Model cost and accuracy are non-linearly correlated, with the cheapest model having the highest accuracy;
• The code generation strategy outperformed direct conversion in almost all scenarios.

1. Large language models have evolved from auxiliary development tools to runtime architectural components, requiring system layers to be redesigned;
2. Dynamic interoperability will become a standard feature of distributed systems to adapt to the dynamic needs of microservices, IoT, and edge computing;
3. Cost optimization is a core consideration for AI-native architectures; models should be selected based on specific needs rather than blindly pursuing expensive commercial models.

SAGAI-MID represents a paradigm shift in the field of distributed system integration, proving the feasibility of dynamic intelligent interoperability in terms of cost and reliability. Paper link: http://arxiv.org/abs/2603.28731v1