[comp] Production Deploy

apps/api/src/vector-store/lib/core/find-similar.spec.ts

@@ -0,0 +1,108 @@
+jest.mock('./client', () => ({
+  vectorIndex: { query: jest.fn() },
+}));
+jest.mock('./generate-embedding', () => ({
+  generateEmbedding: jest.fn(),
+  batchGenerateEmbeddings: jest.fn(),
+}));
+
+import { vectorIndex } from './client';
+import {
+  generateEmbedding,
+  batchGenerateEmbeddings,
+} from './generate-embedding';
+import { findSimilarContent, findSimilarContentBatch } from './find-similar';
+
+const mockQuery = vectorIndex!.query as jest.Mock;
+const mockEmbed = generateEmbedding as jest.Mock;
+const mockBatchEmbed = batchGenerateEmbeddings as jest.Mock;
+
+/**
+ * Builds `count` Upstash results, each from a DISTINCT policy, with strictly
+ * descending scores starting at 0.9. This mirrors CS-594: a single question
+ * matching a chunk of nearly every published policy just above the 0.2 noise
+ * floor.
+ */
+function buildDistinctPolicyResults(count: number) {
+  return Array.from({ length: count }, (_, i) => ({
+    id: `vec_${i}`,
+    score: Number((0.9 - i * 0.02).toFixed(4)),
+    metadata: {
+      content: `Policy chunk ${i}`,
+      sourceType: 'policy',
+      sourceId: `pol_${i}`,
+      policyName: `Policy ${i}`,
+      organizationId: 'org_test',
+    },
+  }));
+}
+
+beforeEach(() => {
+  jest.clearAllMocks();
+  mockEmbed.mockResolvedValue([0.1, 0.2, 0.3]);
+  mockBatchEmbed.mockResolvedValue([[0.1, 0.2, 0.3]]);
+});
+
+describe('findSimilarContent: result cap (CS-594)', () => {
+  it('caps the number of returned chunks even when many distinct policies clear the noise floor', async () => {
+    // 24 distinct policies all scoring >= 0.2 — reproduces Q#17 (24 sources).
+    const flood = buildDistinctPolicyResults(24);
+    expect(flood.every((r) => r.score >= 0.2)).toBe(true);
+    mockQuery.mockResolvedValue(flood);
+
+    const results = await findSimilarContent(
+      'Where is the business located?',
+      'org_test',
+    );
+
+    // Bug: returned all 24. Fix: bounded to a small, relevant set.
+    expect(results.length).toBeLessThan(flood.length);
+    expect(results.length).toBeLessThanOrEqual(10);
+    expect(results.length).toBeGreaterThan(0);
+  });
+
+  it('keeps the highest-scoring chunks and drops the low-scoring tail', async () => {
+    const flood = buildDistinctPolicyResults(24);
+    mockQuery.mockResolvedValue(flood);
+
+    const results = await findSimilarContent('any question', 'org_test');
+
+    // The most relevant chunk must be present, sorted highest-first.
+    expect(results[0].score).toBe(0.9);
+    const scores = results.map((r) => r.score);
+    expect([...scores].sort((a, b) => b - a)).toEqual(scores);
+    // The lowest-scoring chunks of the flood must be dropped, not surfaced.
+    const droppedScore = flood[flood.length - 1].score;
+    expect(scores).not.toContain(droppedScore);
+  });
+
+  it('still filters out chunks below the minimum similarity score', async () => {
+    mockQuery.mockResolvedValue([
+      { id: 'good', score: 0.8, metadata: { sourceType: 'policy' } },
+      { id: 'noise', score: 0.1, metadata: { sourceType: 'policy' } },
+    ]);
+
+    const results = await findSimilarContent('q', 'org_test');
+
+    expect(results.map((r) => r.id)).toEqual(['good']);
+  });
+});
+
+describe('findSimilarContentBatch: result cap (CS-594)', () => {
+  it('caps each question to a small, relevant set instead of every above-threshold chunk', async () => {
+    const flood = buildDistinctPolicyResults(24);
+    mockBatchEmbed.mockResolvedValue([[0.1, 0.2, 0.3]]);
+    mockQuery.mockResolvedValue(flood);
+
+    const [perQuestion] = await findSimilarContentBatch(
+      ['Where is the business located?'],
+      'org_test',
+    );
+
+    expect(perQuestion.length).toBeLessThan(flood.length);
+    expect(perQuestion.length).toBeLessThanOrEqual(10);
+    expect(perQuestion.length).toBeGreaterThan(0);
+    // Highest-scoring chunk preserved.
+    expect(perQuestion[0].score).toBe(0.9);
+  });
+});

apps/api/src/vector-store/lib/core/find-similar.ts

@@ -30,12 +30,20 @@ const MIN_SIMILARITY_SCORE = 0.2;
 // Maximum results to fetch from Upstash Vector (their limit is 1000, but 100 is practical)
 const MAX_TOP_K = 100;
 
+// Maximum distinct chunks returned to the answer generator. Upstash hands back up
+// to MAX_TOP_K candidates above the 0.2 noise floor, but feeding all of them to the
+// LLM floods "Show Sources" with every loosely matching policy (21-24 per question,
+// CS-594) and dilutes the prompt. Keep only the highest-scoring chunks. Mirrors the
+// app-side findSimilarContent limit used by the same auto-answer flow.
+const MAX_RESULTS = 5;
+
 /**
  * Finds similar content using semantic search in Upstash Vector
  * Optimized for RAG (Retrieval-Augmented Generation) answer generation
  *
- * Returns ALL results above the similarity threshold - no artificial limit.
- * This ensures the LLM receives all potentially relevant organizational data.
+ * Returns the highest-scoring results above the similarity threshold, capped at
+ * MAX_RESULTS so the LLM receives the most relevant organizational data without
+ * being flooded by every loosely matching chunk.
  *
  * @param question - The question or query text to search for
  * @param organizationId - Filter results to this organization only
@@ -67,10 +75,12 @@ export async function findSimilarContent(
       filter: `organizationId = "${organizationId}"`,
     });
 
-    // Filter by minimum similarity score only - no artificial limit
-    // All relevant organizational data should reach the LLM
+    // Filter by minimum similarity score, then keep only the highest-scoring
+    // chunks (results are returned sorted by score descending). Without this cap
+    // every chunk above the 0.2 noise floor reaches the LLM and "Show Sources".
     const filteredResults: SimilarContentResult[] = results
       .filter((result) => result.score >= MIN_SIMILARITY_SCORE)
+      .slice(0, MAX_RESULTS)
       .map((result): SimilarContentResult => {
         const metadata = result.metadata as any;
         return {
@@ -193,6 +203,7 @@ export async function findSimilarContentBatch(
     for (const { index, results } of queryResults) {
       const filtered = results
         .filter((result) => result.score >= MIN_SIMILARITY_SCORE)
+        .slice(0, MAX_RESULTS)
         .map((result): SimilarContentResult => {
           const metadata = result.metadata as any;
           return {