https://www.chenk.top/en/tags/json-schema/Recent content in Json-Schema on Chen Kai BlogHugoenThu, 02 Apr 2026 09:00:00 +0000https://www.chenk.top/en/llm-engineering/07-function-calling/Thu, 02 Apr 2026 09:00:00 +0000https://www.chenk.top/en/llm-engineering/07-function-calling/<p>Function calling connects an LLM to the world outside its weights. It combines chat-template details (<a href="https://www.chenk.top/en/llm-engineering/02-tokenization/">Chapter 2</a> ), structured-output kernels (<a href="https://www.chenk.top/en/llm-engineering/05-inference/">Chapter 5</a> ), and prompt engineering (<a href="https://www.chenk.top/en/llm-engineering/09-prompting/">Chapter 9</a> ). This chapter explores what happens under the hood, the guarantees you can rely on, and the agent-loop patterns that handle real workloads.</p> <p>The intellectual lineage matters. Tool use as an LLM capability traces back to two near-simultaneous papers in 2022: <strong>MRKL Systems</strong> (Karpas et al., AI21) which proposed expert-routing among neuro-symbolic modules, and <strong>ReAct</strong> (<a href="https://arxiv.org/abs/2210.03629" target="_blank" rel="noopener noreferrer">Yao et al., 2022 <span aria-hidden="true" style="font-size:0.75em; opacity:0.55; margin-left:2px;">↗</span></a> ) which interleaved chain-of-thought reasoning with tool actions. <strong>Toolformer</strong> (<a href="https://arxiv.org/abs/2302.04761" target="_blank" rel="noopener noreferrer">Schick et al., 2023 <span aria-hidden="true" style="font-size:0.75em; opacity:0.55; margin-left:2px;">↗</span></a> ) showed self-supervised teaching of tool use, generating training data by having a model insert tool-call markers into existing text. By 2024 every frontier model had post-training data structured around the tool-use format, and tool calling moved from &ldquo;research demo&rdquo; to &ldquo;API feature.&rdquo;</p>