arcAgi.ts

import { Effect, pipe } from "effect";
import { genericGeni, provideChatGPT } from ".";
import { Grid } from "./arcAgiDataLoader";
import { getTask } from "./db";
import { BunFileSystem } from "@effect/platform-bun";
import _ from "lodash";

// arcGeni will solve the Arc AGI task by generating a function that matches the description.
const arcGeni = (
  userInstruction: string,
  taskId: string,
  dataset: "training" | "evaluation"
) => {
  const program = Effect.gen(function* () {
    const task = yield* Effect.promise(() => getTask(taskId, dataset));
    if (!task.train || !task.test) {
      throw new Error("No task found");
    }
    const description = `${userInstruction}
    The input is a grid of numbers, and the output is another grid of numbers represented in 2D arrays. Here are the list of {input, output} pair examples:
    ${JSON.stringify(task.train)}`;
    console.log("Calling with description: ", description);
    const fun = yield* genericGeni(
      description,
      [Grid],
      Grid,
      task.train.map(({ input, output }) => ({ input: [input], output }))
    );
    for (const test of task.test) {
      const result = fun(test.input);
      if (!_.isEqual(test.output, result)) {
        throw new Error(`Expected ${test.output} but got ${result}`);
      }
    }
    return fun;
  });

  return Effect.runPromise(
    pipe(program, provideChatGPT, Effect.provide(BunFileSystem.layer))
  );
};

const testing = await arcGeni(
  "Step 1: Identify the pattern of the colored points. Step 2: From the most bottom right blue point, start adding red (2) points from the position where the next blue (1) point would have been, continuing the same pattern. Step 3: Continue adding red points until the next point is out of bounds.",
  "0b17323b",
  "evaluation"
);

console.log(testing);