FIXING TOOL-CALL ACCURACY IN VLLM

Summary from the TLDR Dev newsletter

Ubicloud found that Kimi K2.6 served through vLLM produced schema-valid tool calls only 92% of the time. Constrained decoding raised schema accuracy to 100% for its supported Kimi and DeepSeek models by blocking invalid tokens during generation, though that guarantee does not establish whether a tool call is useful or correct.