Spokes.wiki Search About
Created
Updated

byLLM

The shipped implementation of Meaning-Typed Programming — "Prompt less, Smile more." byLLM is the jac-byllm module inside jaseci's jac monorepo (github.com/jaseci-labs/jac, MIT, 686★/382 forks, created 2021-10-29): the runtime that makes the by language construct actually call an LLM. It's a Jac-language feature first, but the docs state it also ships as a standalone Python library — usable without adopting the rest of Jac.

What it does

Turn a plain function signature into an LLM call by appending by <model>, with no prompt string written anywhere:

def calc_age(birth: date, today: date) -> int by gpt4()

byLLM infers the prompt from the function's name, parameter names, types, and any docstring/type annotations, sends it to the configured model, and parses the response back into the declared return type — including custom types and enums, with no manual output parsing. The site's own framing: "no prompt engineering... a simple by keyword, all the complexity of prompt engineering and output parsing is hidden."

Model-agnostic

Works against OpenAI, Anthropic, Google, or self-hosted models; switching providers is a one-line change. This is the same "capability behind one abstraction, swap the model freely" shape as agent-framework's comparison axis, applied at the language-construct level rather than the library-API level.

Evidence

The accuracy, code-reduction, and cost claims aren't self-reported marketing copy — they're the peer-reviewed measurements in meaning-typed-programming (Dantanarayana et al., PACMPL OOPSLA2 2025), the paper this module implements. See that page for the numbers against DSPy and LMQL.

Shipped examples

The docs point to four real projects built on byLLM: an LLM-driven RPG level generator, an agentic chatbot wired to MCP, a coding agent forked from Aider (aider.chat) with every prompt replaced by a byLLM call (cutting "more than 200 lines of code"), and Jac-GPT, an agentic chatbot running on the full Jaseci stack.

meaning-typed-programming · jaseci · agent-skills · agentic-coding-harness