@rubytech/create-maxy-code 0.1.151 → 0.1.153

package/payload/platform/lib/graph-search/src/__tests__/query-expansion.test.ts

@@ -1,96 +0,0 @@
-import { afterEach, beforeEach, describe, expect, it, vi } from "vitest";
-
-const callMock = vi.fn();
-
-vi.mock("../../../oauth-llm/dist/index.js", () => ({
-  callOauthLlm: (...args: unknown[]) => callMock(...args),
-}));
-
-import { clearExpansionCache, expandQuery } from "../query-expansion.js";
-
-describe("expandQuery", () => {
-  beforeEach(() => {
-    callMock.mockReset();
-    clearExpansionCache();
-  });
-
-  afterEach(() => {
-    callMock.mockReset();
-    clearExpansionCache();
-  });
-
-  it("returns [original, ...paraphrases] on Haiku success", async () => {
-    callMock.mockResolvedValueOnce({
-      kind: "ok-tool",
-      toolName: "emit_paraphrases",
-      input: {
-        paraphrases: ["who is adam langley", "adam langley bio", "adam langley founder"],
-      },
-    });
-    const out = await expandQuery({ query: "Adam Langley", accountId: "acc-1" });
-    expect(out[0]).toBe("Adam Langley");
-    expect(out.slice(1)).toEqual([
-      "who is adam langley",
-      "adam langley bio",
-      "adam langley founder",
-    ]);
-  });
-
-  it("serves the second identical call from cache (no second LLM call)", async () => {
-    callMock.mockResolvedValueOnce({
-      kind: "ok-tool",
-      toolName: "emit_paraphrases",
-      input: { paraphrases: ["p1", "p2", "p3"] },
-    });
-    const first = await expandQuery({ query: "X", accountId: "a" });
-    const second = await expandQuery({ query: "X", accountId: "a" });
-    expect(first).toEqual(second);
-    expect(callMock).toHaveBeenCalledTimes(1);
-  });
-
-  it("caches per retrievalClass — different class re-invokes Haiku", async () => {
-    callMock.mockResolvedValueOnce({
-      kind: "ok-tool",
-      toolName: "emit_paraphrases",
-      input: { paraphrases: ["p1", "p2", "p3"] },
-    });
-    callMock.mockResolvedValueOnce({
-      kind: "ok-tool",
-      toolName: "emit_paraphrases",
-      input: { paraphrases: ["q1", "q2", "q3"] },
-    });
-    const entityCall = await expandQuery({
-      query: "X",
-      accountId: "a",
-      retrievalClass: "entity",
-    });
-    const temporalCall = await expandQuery({
-      query: "X",
-      accountId: "a",
-      retrievalClass: "temporal",
-    });
-    expect(entityCall.slice(1)).toEqual(["p1", "p2", "p3"]);
-    expect(temporalCall.slice(1)).toEqual(["q1", "q2", "q3"]);
-    expect(callMock).toHaveBeenCalledTimes(2);
-  });
-
-  it("falls back to [query] on LLM fallback result", async () => {
-    callMock.mockResolvedValueOnce({
-      kind: "fallback",
-      cause: "missing-creds",
-      reason: "no oauth",
-    });
-    const out = await expandQuery({ query: "X", accountId: "a", useCache: false });
-    expect(out).toEqual(["X"]);
-  });
-
-  it("falls back to [query] when the tool input is malformed", async () => {
-    callMock.mockResolvedValueOnce({
-      kind: "ok-tool",
-      toolName: "emit_paraphrases",
-      input: { paraphrases: "not an array" },
-    });
-    const out = await expandQuery({ query: "X", accountId: "a", useCache: false });
-    expect(out).toEqual(["X"]);
-  });
-});

package/payload/platform/lib/graph-search/src/query-expansion.ts

@@ -1,129 +0,0 @@
-/**
- * Multi-query expansion via Haiku (Task 308).
- *
- * Generates 3-5 paraphrases of the user query so the hybrid pipeline runs
- * each paraphrase against vector + BM25 in parallel and unions the
- * candidate sets before fusion. Recovers recall on terse queries where the
- * embedding's nearest-neighbour is unstable.
- *
- * The expansion runs through `@rubytech/oauth-llm` against Claude Haiku —
- * the same OAuth-bearer path the admin classifiers use; no new SDK
- * dependency and no Anthropic-API-key path. Cache is in-process, keyed by
- * (accountId, normalisedQuery, retrievalClass), with a 1-hour TTL.
- *
- * Graceful degrade: if the LLM call fails, returns just the original
- * query. The caller's hybrid pipeline still runs — just on one query
- * instead of N — so missing paraphrases are a recall regression, not a
- * fault.
- */
-
-import { callOauthLlm } from "../../oauth-llm/dist/index.js";
-
-const CACHE_TTL_MS = 60 * 60 * 1000;
-const HAIKU_MODEL = "claude-haiku-4-5";
-const MAX_PARAPHRASES = 5;
-
-interface CacheEntry {
-  expansions: string[];
-  expiresAt: number;
-}
-
-const cache = new Map<string, CacheEntry>();
-
-export function clearExpansionCache(): void {
-  cache.clear();
-}
-
-export interface ExpandQueryParams {
-  query: string;
-  retrievalClass?: string;
-  accountId?: string;
-  /**
-   * Override the cache or skip it entirely (useful in tests). When
-   * `false`, the cache is read but never written; when `true` (default),
-   * both reads and writes are active.
-   */
-  useCache?: boolean;
-}
-
-/**
- * Return `[query, ...paraphrases]` (original always first). On LLM
- * failure returns `[query]` only and writes a single stderr line so the
- * operator can see expansion is degrading without breaking search.
- */
-export async function expandQuery(
-  params: ExpandQueryParams,
-): Promise<string[]> {
-  const { query, retrievalClass = "general", accountId = "", useCache = true } = params;
-  const normalised = query.trim().toLowerCase();
-  const cacheKey = `${accountId}::${retrievalClass}::${normalised}`;
-
-  if (useCache) {
-    const hit = cache.get(cacheKey);
-    if (hit && hit.expiresAt > Date.now()) {
-      return [query, ...hit.expansions];
-    }
-  }
-
-  const systemPrompt = [
-    "You generate paraphrases of a user query for hybrid graph search.",
-    "Return 3-5 alternative phrasings that preserve the user's intent but vary the surface form (synonyms, word order, entity-vs-attribute framing).",
-    "Do not change the entity names. Do not invent constraints the user did not state.",
-    "Output ONLY a JSON object via the provided tool; never free text.",
-  ].join("\n");
-
-  const classHint = retrievalClass === "general"
-    ? ""
-    : `The retrieval class is '${retrievalClass}' — tailor paraphrases accordingly.`;
-  const userMessage = [
-    `Original query: ${query}`,
-    classHint,
-  ].filter((s) => s.length > 0).join("\n");
-
-  const result = await callOauthLlm({
-    model: HAIKU_MODEL,
-    system: systemPrompt,
-    userMessage,
-    maxTokens: 512,
-    timeoutMs: 15_000,
-    tools: [
-      {
-        name: "emit_paraphrases",
-        description: "Emit 3-5 paraphrases of the user query.",
-        input_schema: {
-          type: "object",
-          properties: {
-            paraphrases: {
-              type: "array",
-              items: { type: "string" },
-              minItems: 3,
-              maxItems: MAX_PARAPHRASES,
-            },
-          },
-          required: ["paraphrases"],
-        },
-      },
-    ],
-    toolChoiceName: "emit_paraphrases",
-  });
-
-  if (result.kind !== "ok-tool") {
-    process.stderr.write(
-      `[graph-search:expansion-failed] cause=${result.kind === "fallback" ? result.cause : "unexpected"} query=${query.slice(0, 60).replace(/\s+/g, " ")}\n`,
-    );
-    return [query];
-  }
-
-  const raw = (result.input as { paraphrases?: unknown }).paraphrases;
-  if (!Array.isArray(raw)) return [query];
-  const expansions = raw
-    .filter((p): p is string => typeof p === "string" && p.trim().length > 0)
-    .map((p) => p.trim())
-    .slice(0, MAX_PARAPHRASES);
-
-  if (useCache) {
-    cache.set(cacheKey, { expansions, expiresAt: Date.now() + CACHE_TTL_MS });
-  }
-
-  return [query, ...expansions];
-}