TurboAgents drops with compressed retrieval, reranking

TurboAgents addresses the "context bottleneck" by implementing 4-bit quantization directly at the retrieval and reranking layers.

• –Employs Walsh-Hadamard rotation and PolarQuant encoding to achieve high-fidelity compressed scoring
• –Validated adapters for Chroma, FAISS, LanceDB, and pgvector allow for "drop-in" integration with existing vector stacks
• –Optimized for a 3.5-bit sweet spot that balances quality and performance for 3B-7B parameter models
• –Framework-agnostic design enables compressed retrieval without requiring a full infrastructure migration
• –Essential for local agent systems where VRAM and memory-to-compute ratios are the primary constraints