fix(search): cap vector hydration window

convex/search.test.ts

@@ -1060,17 +1060,18 @@ describe("search helpers", () => {
     expect(result[0].skill.slug).toBe("fallback-skill");
   });
 
-  it("hydrates the stable max vector window for ordinary load-more searches", async () => {
+  it("hydrates a bounded vector window for ordinary load-more searches", async () => {
     generateEmbeddingMock.mockResolvedValueOnce([0, 1, 2]);
 
-    // Ordinary first-page and load-more searches use a stable recall floor, so
-    // candidateLimit starts at the Convex vector maximum.
-    const batch = Array.from({ length: 256 }, (_, i) => ({
+    const batch = Array.from({ length: 128 }, (_, i) => ({
       _id: `skillEmbeddings:e${i}`,
       _score: 0.5 - i * 0.001,
     }));
 
-    const vectorSearchMock = vi.fn().mockResolvedValueOnce(batch);
+    const vectorSearchMock = vi.fn(
+      async (_table: unknown, _index: unknown, opts: { limit: number }) =>
+        batch.slice(0, opts.limit),
+    );
 
     const hydrateCalls: string[][] = [];
     const runQuery = vi.fn(
@@ -1101,9 +1102,10 @@ describe("search helpers", () => {
       { query: "test", limit: 50 },
     );
 
-    expect(vectorSearchMock).toHaveBeenCalledTimes(1);
-    expect(hydrateCalls).toHaveLength(1);
-    expect(hydrateCalls[0]).toHaveLength(256);
+    expect(vectorSearchMock).toHaveBeenCalledTimes(2);
+    expect(hydrateCalls).toHaveLength(2);
+    expect(hydrateCalls[0]).toHaveLength(100);
+    expect(hydrateCalls[1]).toHaveLength(28);
   });
 
   it("merges fallback matches without duplicate skill ids", () => {
@@ -1239,7 +1241,7 @@ describe("soul search", () => {
 
   it("hydrates only new soul embedding ids across vector iterations", async () => {
     generateEmbeddingMock.mockResolvedValueOnce([0, 1, 2]);
-    const firstBatch = Array.from({ length: 200 }, (_, i) => ({
+    const firstBatch = Array.from({ length: 100 }, (_, i) => ({
       _id: i === 0 ? "soulEmbeddings:a" : `soulEmbeddings:filler${i}`,
       _score: i === 0 ? 0.9 : 0.1,
     }));

convex/search.ts

@@ -60,6 +60,8 @@ const POPULARITY_WEIGHT = 0.08;
 const FALLBACK_SCAN_LIMIT = 2000;
 const MIN_STABLE_SEARCH_RECALL_LIMIT = 100;
 const MAX_DIRECT_SKILL_SEARCH_CANDIDATES = 100;
+const MIN_VECTOR_SEARCH_CANDIDATES = 50;
+const MAX_VECTOR_SEARCH_CANDIDATES = 128;
 const SKILL_CAPABILITY_TAG_SET = new Set<string>(SKILL_CAPABILITY_TAGS);
 
 function getNextCandidateLimit(current: number, max: number) {
@@ -188,11 +190,13 @@ export const searchSkills: ReturnType<typeof action> = action({
     // Keep ordinary first-page and load-more requests ranking the same recall pool
     // before slicing, so expanding the display limit does not reshuffle the prefix.
     const recallLimit = Math.max(limit, MIN_STABLE_SEARCH_RECALL_LIMIT);
-    // Convex vectorSearch max limit is 256; clamp candidate sizes accordingly.
-    // Keep the initial pool large enough to catch moderate-vector matches
-    // that win after lexical and popularity scoring, even for small limits.
-    const maxCandidate = Math.min(Math.max(recallLimit * 10, 200), 256);
-    let candidateLimit = Math.min(Math.max(recallLimit * 3, 200), 256);
+    // Keep the vector pool bounded; exact slug, prefix, and lexical fallback cover
+    // literal recall without hydrating hundreds of semantic candidates per search.
+    const maxCandidate = Math.min(
+      Math.max(limit * 4, MIN_VECTOR_SEARCH_CANDIDATES),
+      MAX_VECTOR_SEARCH_CANDIDATES,
+    );
+    let candidateLimit = Math.min(Math.max(limit * 2, MIN_VECTOR_SEARCH_CANDIDATES), maxCandidate);
     let hydrated: SkillSearchEntry[] = [];
     const seenEmbeddingIds = new Set<Id<"skillEmbeddings">>();
     let scoreById = new Map<Id<"skillEmbeddings">, number>();
@@ -664,10 +668,11 @@ export const searchSouls: ReturnType<typeof action> = action({
       vector = null;
     }
     const limit = args.limit ?? 10;
-    // Convex vectorSearch max limit is 256; clamp candidate sizes accordingly.
-    // Match searchSkills so soul search does not miss boosted exact matches.
-    const maxCandidate = Math.min(Math.max(limit * 10, 200), 256);
-    let candidateLimit = Math.min(Math.max(limit * 3, 200), 256);
+    const maxCandidate = Math.min(
+      Math.max(limit * 4, MIN_VECTOR_SEARCH_CANDIDATES),
+      MAX_VECTOR_SEARCH_CANDIDATES,
+    );
+    let candidateLimit = Math.min(Math.max(limit * 2, MIN_VECTOR_SEARCH_CANDIDATES), maxCandidate);
     let hydrated: HydratedSoulEntry[] = [];
     const seenEmbeddingIds = new Set<Id<"soulEmbeddings">>();
     let scoreById = new Map<Id<"soulEmbeddings">, number>();