chore(search): retire legacy search knowledge base projection rows in place

apps/sim/lib/knowledge/__integration__/workspace-kb-document-access.integration.ts

@@ -6,7 +6,7 @@
  * nothing about what a workspace search returns. A signed-in reader's own grants widen what they
  * read, and a source whose reader must be proven live admits a candidate only once the proof
  * holds, with the readable rows below it filling the page when it does not. A Search-marked knowledge base
- * named explicitly by id is searched the same way.
+ * named explicitly by id is searched the same way, except that it keeps no vector candidates.
  */
 import { createHash } from 'node:crypto'
 import type { Principal } from '@sim/auth/principal'
@@ -48,10 +48,14 @@ afterAll(async () => {
   await db.$client.end()
 })
 
+const ALL_MODES: readonly KnowledgeSearchMode[] = ['vector', 'hybrid']
+/** A Search knowledge base's chunk writes delete its candidate rows, so only keyword ranking finds it. */
+const KEYWORD_RANKED_MODES: readonly KnowledgeSearchMode[] = ['hybrid']
+
 describe.each([
-  ['a workspace knowledge base', false],
-  ['a dormant search index', true],
-] as const)('search of %s decides access on the document', (_label, isSearchIndex) => {
+  ['a workspace knowledge base', false, ALL_MODES],
+  ['a dormant search index', true, KEYWORD_RANKED_MODES],
+] as const)('search of %s decides access on the document', (_label, isSearchIndex, modes) => {
   const ids = createKnowledgeAclFixtureIds()
   const baseId = generateId()
   const readable = [generateId(), generateId(), generateId()]
@@ -161,41 +165,43 @@ describe.each([
     return result.rows.map((row) => row.documentId).sort()
   }
 
-  it.each(['vector', 'hybrid'] as const)(
-    'returns a signed-in reader the rows an actorless run gets (%s)',
-    async (mode) => {
-      const previousDebug = db.$client.options.debug
-      const statements: string[] = []
-      db.$client.options.debug = (_connection, query) => {
-        statements.push(query)
-      }
-      try {
-        const signedIn = await search(reader, mode)
-        const workspaceKey = await search(actorless, mode)
-        expect(signedIn).toEqual([...readable].sort())
-        expect(workspaceKey).toEqual(signedIn)
-        /** Nothing read the per-row machinery the search-index path keeps for itself. */
-        for (const fragment of [
-          'embedding_keyword_search',
-          'knowledge_projection_dirty',
-          '"embedding_search"."acl"',
-          '"embedding_search"."connector_id"',
-        ]) {
-          expect(statements.filter((statement) => statement.includes(fragment))).toEqual([])
-        }
-      } finally {
-        db.$client.options.debug = previousDebug
+  it.each(modes)('returns a signed-in reader the rows an actorless run gets (%s)', async (mode) => {
+    const previousDebug = db.$client.options.debug
+    const statements: string[] = []
+    db.$client.options.debug = (_connection, query) => {
+      statements.push(query)
+    }
+    try {
+      const signedIn = await search(reader, mode)
+      const workspaceKey = await search(actorless, mode)
+      expect(signedIn).toEqual([...readable].sort())
+      expect(workspaceKey).toEqual(signedIn)
+      /** Nothing read the per-row machinery the search-index path keeps for itself. */
+      for (const fragment of [
+        'embedding_keyword_search',
+        'knowledge_projection_dirty',
+        '"embedding_search"."acl"',
+        '"embedding_search"."connector_id"',
+      ]) {
+        expect(statements.filter((statement) => statement.includes(fragment))).toEqual([])
       }
+    } finally {
+      db.$client.options.debug = previousDebug
     }
-  )
+  })
 
-  it.each(['vector', 'hybrid'] as const)(
+  it.each(modes)(
     'returns a signed-in reader the document shared with them alone as well (%s)',
     async (mode) => {
       const owner: Principal = { kind: 'session', userId: ids.aliceId, sessionId: 'fixture-owner' }
       expect(await search(owner, mode)).toEqual([...readable, personal].sort())
     }
   )
+
+  it.runIf(isSearchIndex)('finds nothing by vector, since no candidate row is kept', async () => {
+    const owner: Principal = { kind: 'session', userId: ids.aliceId, sessionId: 'fixture-owner' }
+    expect(await search(owner, 'vector')).toEqual([])
+  })
 })
 
 describe('a workspace source whose reader must be proven live', () => {

packages/db/knowledge-projection.ts

@@ -68,11 +68,16 @@ const SEARCH_VECTOR_COLUMNS = SEARCH_VECTOR_WIDTHS.map((width) => widthColumn('v
 /** The bit columns of `embedding_search`, whose width check requires exactly one. */
 const SEARCH_BINARY_COLUMNS = ['"binary"', 'binary_384', 'binary_768', 'binary_1024', 'binary_3072']
 
+/** The `ON CONFLICT DO UPDATE` assignments that take each column from the proposed row. */
+function excludedAssignments(columns: readonly string[]): string {
+  return columns.map((column) => `${column} = EXCLUDED.${column}`).join(', ')
+}
+
 /**
  * The halfvec projections of an embedding row, in {@link SEARCH_VECTOR_COLUMNS} order. Shortening
  * is valid only for the two OpenAI models trained for prefix retrieval. These and the bit
- * projections match what `sync_embedding_search()` from `0016_backfill_search_vectors` writes, so
- * a pass over rows a synchronous writer wrote finds nothing to rewrite.
+ * projections are what {@link SYNC_EMBEDDING_SEARCH_FUNCTION} writes, so a pass over rows a
+ * synchronous writer wrote finds nothing to rewrite.
  */
 function searchVectorProjections(prefix: string, shortened: string): string {
   return SEARCH_VECTOR_WIDTHS.map((width) =>
@@ -94,6 +99,51 @@ function searchVectorShortened(model: string, prefix: string): string {
   return `${model} IN ${SHORTENED_EMBEDDING_MODELS} AND ${prefix}.embedding_384 IS NULL`
 }
 
+/**
+ * The `embedding_search_sync` trigger function: projects an ordinary-KB chunk and deletes the row
+ * of a Search-KB chunk. Migration 0449 installs this exact body; fresh databases and pushes install
+ * it through `0016_backfill_search_vectors`.
+ *
+ * contract-pending(after Search KB candidate rows are purged and Search KBs are unwritable): drop
+ * the `search_index` delete branch — no Search-KB chunk write can reach the trigger then.
+ */
+export const SYNC_EMBEDDING_SEARCH_FUNCTION = `CREATE OR REPLACE FUNCTION sync_embedding_search()
+  RETURNS trigger LANGUAGE plpgsql AS $$
+  DECLARE
+    search_index boolean;
+    shortened boolean;
+  BEGIN
+    SELECT is_search_index, ${searchVectorShortened('embedding_model', 'NEW')}
+      INTO STRICT search_index, shortened
+    FROM knowledge_base WHERE id = NEW.knowledge_base_id;
+    IF search_index THEN
+      DELETE FROM embedding_search WHERE id = NEW.id;
+      RETURN NEW;
+    END IF;
+    INSERT INTO embedding_search
+      (id, knowledge_base_id, document_id, enabled, ${SEARCH_BINARY_COLUMNS.join(', ')},
+        ${SEARCH_VECTOR_COLUMNS.join(', ')})
+    VALUES (NEW.id, NEW.knowledge_base_id, NEW.document_id, NEW.enabled,
+      ${searchBinaryProjections('NEW')},
+      ${searchVectorProjections('NEW', 'shortened')})
+    ON CONFLICT (id) DO UPDATE SET
+      ${excludedAssignments(['knowledge_base_id', 'document_id', 'enabled', ...SEARCH_BINARY_COLUMNS, ...SEARCH_VECTOR_COLUMNS])};
+    RETURN NEW;
+  END;
+  $$`
+
+const comparableSource = (source: string) => source.replace(/\s+/g, ' ').trim()
+
+/**
+ * Whether a live `pg_proc.prosrc` is {@link SYNC_EMBEDDING_SEARCH_FUNCTION}'s body, ignoring
+ * whitespace: PostgreSQL stores the body between the `$$` quotes verbatim.
+ */
+export function isCurrentSyncEmbeddingSearchSource(prosrc: string): boolean {
+  return (
+    comparableSource(prosrc) === comparableSource(SYNC_EMBEDDING_SEARCH_FUNCTION.split('$$')[1])
+  )
+}
+
 export type KnowledgeProjection = 'embedding_search'
 
 /** Active vector projections that mirror document access for legacy synchronous writers. */
@@ -138,7 +188,7 @@ function contentPageStatement(): string {
         ${searchVectorProjections('source', 'shortened')}
       FROM source
       ON CONFLICT (id) DO UPDATE SET
-        ${[...compared, ...SEARCH_BINARY_COLUMNS].map((column) => `${column} = EXCLUDED.${column}`).join(', ')}
+        ${excludedAssignments([...compared, ...SEARCH_BINARY_COLUMNS])}
       WHERE (${compared.map((column) => `s.${column}`).join(', ')})
         IS DISTINCT FROM (${compared.map((column) => `EXCLUDED.${column}`).join(', ')})
       RETURNING s.id