Asynchronous LLM computations specifications with LLM:Graph (rakuforprediction.wordpress.com)

Anton Antonov released LLM::Graph, a Raku package that models multi-step LLM workflows as directed computation graphs to schedule and combine multiple LLM generation steps efficiently. It’s installable via zef (zef install LLM::Graph) or directly from GitHub, and requires normal LLM service authentication and internet access for LLM nodes. LLM::Graph instances are callable objects (you can call $g(input) or $g.eval(input)), expose a default llm-evaluator, and support async execution via a boolean async option that uses Promise to submit concurrent LLM requests. The author demonstrates patterns (e.g., three poet nodes and a judge node) and provides notebooks for visualization and interpretation. Technically, nodes are specified as either eval-function (Raku subs), llm-function (LLM submission), or listable-llm-function (threaded LLM on list inputs); each node can declare inputs, conditional test-function logic, and templated prompts. llm-function prompts accept static strings, string lists, repository prompts, templated subs, or LLM::Function objects — and llm-functions now produce function objects (functors) by default. The graph is converted to a Graph object and evaluated via recursion (topological sorting proved unnecessary), with special wrapping for string-template subs (blocks currently unsupported). Visual output includes richer plots with distinct node shapes and dashed edges for test dependencies, making complex LLM pipelines both performant and debuggable.