Replace intent dimensions with corpus-backed problem picker

packages/agent/src/index.ts

@@ -84,6 +84,7 @@ export async function main(): Promise<void> {
 
   const app = createApp({
     mathEngine,
+    rag,
     workflow: {
       rag,
       mathEngine,

packages/agent/src/server/app.ts

@@ -4,12 +4,15 @@ import { Hono } from "hono";
 import { cors } from "hono/cors";
 
 import type { MathEngineClient } from "../tools/math-engine-client.js";
+import type { RagClient } from "../tools/rag-client.js";
 import type { RunOptions, VerificationWorkflowDeps } from "../workflows/verification-workflow.js";
 import { createGenerateRoute } from "./routes/generate.js";
 import { createHealthRoute } from "./routes/health.js";
+import { createSourceProblemsRoute } from "./routes/source-problems.js";
 
 export interface AppDeps {
   mathEngine: MathEngineClient;
+  rag: RagClient;
   workflow: VerificationWorkflowDeps;
   workflowOptions?: RunOptions;
 }
@@ -27,6 +30,7 @@ export function createApp(deps: AppDeps): Hono {
   );
 
   app.route("/", createHealthRoute(deps.mathEngine));
+  app.route("/", createSourceProblemsRoute(deps.rag));
   app.route("/", createGenerateRoute(deps.workflow, deps.workflowOptions));
 
   return app;

packages/agent/src/server/routes/source-problems.ts

@@ -0,0 +1,81 @@
+/**
+ * GET /api/source-problems — read-only corpus lookup.
+ *
+ * Reuses the in-memory RagClient (D-7) without LLM / generation. The web S3
+ * problem picker calls this to browse SourceProblem rows by school/grade/topic.
+ *
+ * Figure-dependent rows are dropped: the corpus image path is a non-servable
+ * AI-Hub scan, so a statement that points at a figure/graph/table is unusable
+ * as text-only. See isFigureDependent for why there is no clean corpus flag.
+ */
+
+import { zValidator } from "@hono/zod-validator";
+import { Hono } from "hono";
+import { z } from "zod";
+
+import {
+  DifficultySchema,
+  SchoolLevelSchema,
+} from "../../schemas/source-problem.schema.js";
+import type { RagClient } from "../../tools/rag-client.js";
+
+// Big k = scan whole topic (≥ largest ~1.8k rows). search() already sorts the full
+// match set so this only widens the final slice; lets figure filtering fill `limit`
+// on figure-heavy topics instead of returning an all-figure first-50.
+const FIGURE_FILTER_POOL = 2000;
+
+// No reliable corpus flag for "needs a figure": media.image is on 100% of rows, and
+// the VISUAL taxonomy tag also marks self-contained solids ("직육면체의 겉넓이"). So match
+// only explicit references — "그림", a shown graph/도형/표 via deictic phrases, or embedded
+// <table>. Bare "그래프"/"도형"/"표"/"좌표" stay (e.g. "y=x²의 그래프의 꼭짓점" is self-contained).
+const FIGURE_REFERENCE =
+  /그림|그래프와\s*같|(?:다음|아래|위)\s*그래프|(?:다음|아래)\s*도형|(?:다음|아래|위)의?\s*표(?:는|를|에|\s)|<table/i;
+
+export function isFigureDependent(questionText: string): boolean {
+  return FIGURE_REFERENCE.test(questionText);
+}
+
+const GradeQueryParamSchema = z.preprocess(
+  (value) => {
+    if (value === "common" || value === "null" || value === "") return null;
+    if (value === "1" || value === 1) return 1;
+    if (value === "2" || value === 2) return 2;
+    if (value === "3" || value === 3) return 3;
+    return value;
+  },
+  z.union([z.literal(1), z.literal(2), z.literal(3)]).nullable(),
+);
+
+const SourceProblemsQuerySchema = z.object({
+  school_level: SchoolLevelSchema,
+  grade: GradeQueryParamSchema,
+  topic_code: z.string().min(1).optional(),
+  difficulty: DifficultySchema.optional(),
+  limit: z.coerce.number().int().min(1).max(50).default(30),
+});
+
+export function createSourceProblemsRoute(rag: RagClient): Hono {
+  const app = new Hono();
+
+  app.get(
+    "/api/source-problems",
+    zValidator("query", SourceProblemsQuerySchema),
+    async (c) => {
+      const q = c.req.valid("query");
+      const results = await rag.search({
+        school_level: q.school_level,
+        grade: q.grade ?? null,
+        topic_code: q.topic_code,
+        difficulty: q.difficulty,
+        k: FIGURE_FILTER_POOL,
+      });
+      const problems = results
+        .map((r) => r.problem)
+        .filter((p) => !isFigureDependent(p.question_text))
+        .slice(0, q.limit);
+      return c.json(problems);
+    },
+  );
+
+  return app;
+}

packages/agent/tests/source-problems-route.test.ts

@@ -0,0 +1,185 @@
+import { Hono } from "hono";
+import { describe, expect, it } from "vitest";
+
+import type {
+  RagQuery,
+  RagResult,
+  SourceProblem,
+} from "../src/schemas/index.js";
+import {
+  createSourceProblemsRoute,
+  isFigureDependent,
+} from "../src/server/routes/source-problems.js";
+import type { RagClient } from "../src/tools/rag-client.js";
+
+function makeProblem(overrides: {
+  item_id: string;
+  difficulty_norm: SourceProblem["difficulty_norm"];
+  topic_code?: SourceProblem["topic_code"];
+  school_level?: SourceProblem["school_level"];
+  grade?: SourceProblem["grade"];
+  question_text?: string;
+}): SourceProblem {
+  return {
+    item_id: overrides.item_id,
+    source_dataset: "111",
+    split: "train",
+    source_label_type: "problem_label",
+    school_level: overrides.school_level ?? "middle",
+    grade: overrides.grade ?? 3,
+    semester: null,
+    topic_code: overrides.topic_code ?? "9수02-09",
+    topic_name: "이차방정식",
+    achievement_standard: "이차방정식을 풀 수 있다.",
+    question_text: overrides.question_text ?? `질문 ${overrides.item_id}`,
+    answer_text: `정답 ${overrides.item_id}`,
+    explanation_text: null,
+    choice_blocks: null,
+    problem_type_norm: "short_answer",
+    difficulty_norm: overrides.difficulty_norm,
+    question_image_relpath: null,
+    answer_image_relpath: null,
+    question_json_relpath: null,
+    answer_json_relpath: null,
+  };
+}
+
+function createMockRag(problems: SourceProblem[]): RagClient {
+  return {
+    async search(query: RagQuery): Promise<RagResult[]> {
+      const filtered = problems.filter((problem) => {
+        if (problem.school_level !== query.school_level) return false;
+        if (query.grade !== null && problem.grade !== query.grade) return false;
+        if (query.topic_code && problem.topic_code !== query.topic_code) {
+          return false;
+        }
+        if (query.difficulty && problem.difficulty_norm !== query.difficulty) {
+          return false;
+        }
+        return true;
+      });
+
+      return filtered.slice(0, query.k ?? 8).map((problem) => ({
+        item_id: problem.item_id,
+        similarity: 0.5,
+        problem,
+        match_reason: "structural",
+      }));
+    },
+  };
+}
+
+describe("GET /api/source-problems", () => {
+  const fixtures: SourceProblem[] = [
+    makeProblem({ item_id: "mock-1", difficulty_norm: "easy" }),
+    makeProblem({ item_id: "mock-2", difficulty_norm: "medium" }),
+    makeProblem({ item_id: "mock-3", difficulty_norm: "hard" }),
+    makeProblem({ item_id: "mock-4", difficulty_norm: "hard" }),
+    makeProblem({
+      item_id: "mock-other-topic",
+      difficulty_norm: "medium",
+      topic_code: "9수02-08",
+    }),
+    makeProblem({
+      item_id: "mock-other-grade",
+      difficulty_norm: "easy",
+      grade: 2,
+    }),
+    makeProblem({
+      item_id: "mock-figure",
+      difficulty_norm: "medium",
+      question_text: "다음 그림과 같이 정육면체를 잘라 낸 입체도형에서 면의 개수를 구하여라.",
+    }),
+  ];
+
+  function buildApp(): Hono {
+    const app = new Hono();
+    app.route("/", createSourceProblemsRoute(createMockRag(fixtures)));
+    return app;
+  }
+
+  it("returns source problems filtered by school_level + grade + topic_code", async () => {
+    const app = buildApp();
+    const topic = encodeURIComponent("9수02-09");
+    const res = await app.request(
+      `/api/source-problems?school_level=middle&grade=3&topic_code=${topic}`,
+    );
+
+    expect(res.status).toBe(200);
+    const body = (await res.json()) as SourceProblem[];
+    expect(Array.isArray(body)).toBe(true);
+    expect(body).toHaveLength(4);
+    for (const item of body) {
+      expect(item).toHaveProperty("item_id");
+      expect(item).toHaveProperty("question_text");
+      expect(item).toHaveProperty("answer_text");
+      expect(item).toHaveProperty("difficulty_norm");
+      expect(item.topic_code).toBe("9수02-09");
+      expect(item.school_level).toBe("middle");
+      expect(item.grade).toBe(3);
+    }
+  });
+
+  it("returns 400 when school_level is missing", async () => {
+    const app = buildApp();
+    const res = await app.request("/api/source-problems?grade=3");
+    expect(res.status).toBe(400);
+  });
+
+  it("narrows results when difficulty=hard", async () => {
+    const app = buildApp();
+    const topic = encodeURIComponent("9수02-09");
+    const res = await app.request(
+      `/api/source-problems?school_level=middle&grade=3&topic_code=${topic}&difficulty=hard`,
+    );
+
+    expect(res.status).toBe(200);
+    const body = (await res.json()) as SourceProblem[];
+    expect(body).toHaveLength(2);
+    expect(body.every((problem) => problem.difficulty_norm === "hard")).toBe(true);
+  });
+
+  it("returns [] for an unknown topic_code", async () => {
+    const app = buildApp();
+    const topic = encodeURIComponent("9수99-99");
+    const res = await app.request(
+      `/api/source-problems?school_level=middle&grade=3&topic_code=${topic}`,
+    );
+
+    expect(res.status).toBe(200);
+    const body = (await res.json()) as SourceProblem[];
+    expect(body).toEqual([]);
+  });
+
+  it("excludes figure-dependent problems (statement references 그림)", async () => {
+    const app = buildApp();
+    const topic = encodeURIComponent("9수02-09");
+    const res = await app.request(
+      `/api/source-problems?school_level=middle&grade=3&topic_code=${topic}`,
+    );
+
+    expect(res.status).toBe(200);
+    const body = (await res.json()) as SourceProblem[];
+    expect(body.some((p) => p.item_id === "mock-figure")).toBe(false);
+    expect(body.every((p) => !p.question_text.includes("그림"))).toBe(true);
+  });
+});
+
+describe("isFigureDependent", () => {
+  it.each([
+    ["다음 그림과 같이 정육면체를 잘라 낸 입체도형에서", true],
+    ["다음 그래프는 x초 후 높이 y를 나타낸 것이다.", true],
+    ["아래 그래프는 시간과 거리의 관계를 나타낸 것이다.", true],
+    ["걸린 시간과 이동 거리를 그래프와 같이 나타내면", true],
+    ["다음 표는 어느 반의 도수분포표이다.", true],
+    ["다음 도형의 둘레의 길이를 구하여라.", true],
+    ["성적을 정리하면 <table border> 와 같다.", true],
+    ["모든 모서리의 합이 48이고 대각선이 3√6인 직육면체의 겉넓이를 구하시오.", false],
+    ["이차함수 y=2x^2+3x-2의 그래프와 x축의 교점의 x좌표를 구하시오.", false],
+    ["수직선 위에서 -7/5보다 큰 가장 작은 정수를 구하시오.", false],
+    ["a^3+b^3을 보기와 같이 변형하여 값을 구하시오.", false],
+    ["이차방정식 (2x+1)(x-4)=0을 풀어라.", false],
+  ])("%s -> %s", (text, expected) => {
+    expect(isFigureDependent(text)).toBe(expected);
+  });
+});