fix(schema): skip unsupported large-dim HNSW indexes (#1734)

Takeover of #2510: migrations v40 (facts) and v55 (query_cache)
unconditionally created HNSW indexes with the configured embedding
dimension, so `gbrain init` with embedding_dimensions above pgvector's
per-type HNSW caps (vector 2000 / halfvec 4000) failed with
"column cannot have more than 4000 dimensions for hnsw index".

- vector-index.ts: add PGVECTOR_HNSW_HALFVEC_MAX_DIMS + hnswMaxDimsForType
- migrate.ts v40/v55: emit the HNSW index only when dims fit the cap,
  otherwise a comment noting exact scans remain available
- embedding-dim-check.ts: buildFactsAlterRecipe skips the reindex step
  above the cap for the same reason
- tests: 4096d init round-trip on PGLite (columns exist, indexes
  skipped) + recipe-skip unit test

Drops the unrelated context-engine.ts interface change and the
tsconfig.json strictFunctionTypes=false hunk from #2510; typecheck is
clean without them.

Co-authored-by: javieraldape <javieraldape@users.noreply.github.com>
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

src/commands/sync.ts

@@ -197,7 +197,7 @@ export interface SyncResult {
   /** Pages re-embedded during this sync's auto-embed step. 0 if --no-embed or skipped. */
   embedded: number;
   pagesAffected: string[];
-  failedFiles?: number; // count of per-file import/sync failures (Bug 9)
+  failedFiles?: number; // count of parse failures (Bug 9)
   /**
    * v0.41.13.0 partial-sync fields (only set when status === 'partial').
    *
@@ -3183,7 +3183,7 @@ async function performSyncInner(engine: BrainEngine, opts: SyncOpts): Promise<Sy
     await clearOpCheckpoint(engine, ckpt.target);
   };
 
-  // issue #1939 adversarial finding #1: a file that failed to import (open ledger
+  // issue #1939 adversarial finding #1: a file that failed to parse (open ledger
   // row) and is then deleted/renamed-away never re-enters failedFiles and never
   // imports, so its row would never clear and would age doctor to a permanent
   // FAIL. Treat removed paths as resolved so the ledger self-heals.
@@ -3215,9 +3215,9 @@ async function performSyncInner(engine: BrainEngine, opts: SyncOpts): Promise<Sy
     } else {
       const fileFailCount = failedFiles.filter(f => isSkippablePath(f.path)).length;
       serr(
-        `\nSync blocked: ${fileFailCount} file(s) failed to import:\n` +
+        `\nSync blocked: ${fileFailCount} file(s) failed to parse:\n` +
         `${codeBreakdown}\n\n` +
-        `Fix the listed file errors and re-run, or use 'gbrain sync --skip-failed' to ` +
+        `Fix the frontmatter and re-run, or use 'gbrain sync --skip-failed' to ` +
         `acknowledge and move on. A file that keeps failing auto-skips after ` +
         `${resolveAutoSkipThreshold()} consecutive syncs.`,
       );
@@ -5355,7 +5355,7 @@ function printSyncResult(result: SyncResult, sink: NodeJS.WriteStream = process.
     case 'dry_run':
       break; // already printed in performSync
     case 'blocked_by_failures':
-      write(`Sync BLOCKED at ${result.toCommit.slice(0, 8)}: ${result.failedFiles ?? 0} file(s) failed to import.`);
+      write(`Sync BLOCKED at ${result.toCommit.slice(0, 8)}: ${result.failedFiles ?? 0} file(s) failed to parse.`);
       write(`  See ~/.gbrain/sync-failures.jsonl for details, or run 'gbrain doctor'.`);
       write(`  Fix the files then re-run 'gbrain sync', or 'gbrain sync --skip-failed' to move on.`);
       break;

src/core/embedding-dim-check.ts

@@ -14,7 +14,7 @@
  */
 
 import type { BrainEngine } from './engine.ts';
-import { PGVECTOR_HNSW_VECTOR_MAX_DIMS } from './vector-index.ts';
+import { PGVECTOR_HNSW_VECTOR_MAX_DIMS, hnswMaxDimsForType } from './vector-index.ts';
 import { gbrainPath } from './config.ts';
 import { resolveRecipe } from './ai/model-resolver.ts';
 import type { Recipe } from './ai/types.ts';
@@ -609,6 +609,17 @@ export function buildFactsAlterRecipe(
   const opclass = columnType === 'halfvec' ? 'halfvec_cosine_ops' : 'vector_cosine_ops';
   const targetType = columnType === 'halfvec' ? `halfvec(${configuredDims})` : `vector(${configuredDims})`;
   const dimsChanged = columnDims !== configuredDims;
+  const hnswMaxDims = hnswMaxDimsForType(columnType);
+  const indexLines = configuredDims <= hnswMaxDims
+    ? [
+        `CREATE INDEX idx_facts_embedding_hnsw`,
+        `  ON facts USING hnsw (embedding ${opclass})`,
+        `  WHERE embedding IS NOT NULL AND expired_at IS NULL;`,
+      ]
+    : [
+        `-- Skip reindex. ${columnType}(${configuredDims}) exceeds pgvector's HNSW cap of ${hnswMaxDims};`,
+        `-- fact similarity falls back to exact scans.`,
+      ];
   return [
     `-- ALTER ${columnType}(${columnDims}) → ${columnType}(${configuredDims}) on indexed column.`,
     `-- HOLD a maintenance window: this rewrites every row's embedding.`,
@@ -629,9 +640,7 @@ export function buildFactsAlterRecipe(
       : []),
     `ALTER TABLE facts ALTER COLUMN embedding TYPE ${targetType}`,
     `  USING embedding::${targetType};`,
-    `CREATE INDEX idx_facts_embedding_hnsw`,
-    `  ON facts USING hnsw (embedding ${opclass})`,
-    `  WHERE embedding IS NOT NULL AND expired_at IS NULL;`,
+    ...indexLines,
   ].join('\n');
 }
 

src/core/migrate.ts

@@ -1,6 +1,7 @@
 import type { BrainEngine } from './engine.ts';
 import { slugifyPath } from './sync.ts';
 import { getFtsLanguage } from './fts-language.ts';
+import { hnswMaxDimsForType } from './vector-index.ts';
 
 /**
  * Schema migrations — run automatically on initSchema().
@@ -2276,11 +2277,19 @@ export const MIGRATIONS: Migration[] = [
         useHalfvec = true;
       }
 
-      const vecType = useHalfvec ? 'HALFVEC' : 'VECTOR';
+      const columnType = useHalfvec ? 'halfvec' : 'vector';
+      const vecType = columnType.toUpperCase();
       // HNSW operator class must match the column type:
       //   VECTOR(n)  → vector_cosine_ops
       //   HALFVEC(n) → halfvec_cosine_ops
       const opclass = useHalfvec ? 'halfvec_cosine_ops' : 'vector_cosine_ops';
+      const hnswMaxDims = hnswMaxDimsForType(columnType);
+      const factsEmbeddingIndexSql = embeddingDim <= hnswMaxDims
+        ? `CREATE INDEX IF NOT EXISTS idx_facts_embedding_hnsw
+          ON facts USING hnsw (embedding ${opclass})
+          WHERE embedding IS NOT NULL AND expired_at IS NULL;`
+        : `-- idx_facts_embedding_hnsw skipped: pgvector HNSW ${columnType} indexes support
+        -- at most ${hnswMaxDims} dimensions; exact vector scans remain available.`;
       // FK to sources is added in a separate ALTER TABLE rather than inline
       // on the column. Inline `REFERENCES` worked on PGLite but silently
       // got dropped by postgres.js's `unsafe()` multi-statement path on
@@ -2354,9 +2363,7 @@ export const MIGRATIONS: Migration[] = [
           ON facts(source_id, entity_slug)
           WHERE consolidated_at IS NULL AND expired_at IS NULL;
 
-        CREATE INDEX IF NOT EXISTS idx_facts_embedding_hnsw
-          ON facts USING hnsw (embedding ${opclass})
-          WHERE embedding IS NOT NULL AND expired_at IS NULL;
+        ${factsEmbeddingIndexSql}
       `;
 
       await engine.runMigration(40, factsDDL);
@@ -2870,8 +2877,16 @@ export const MIGRATIONS: Migration[] = [
         useHalfvec = true;
       }
 
-      const vecType = useHalfvec ? 'HALFVEC' : 'VECTOR';
+      const columnType = useHalfvec ? 'halfvec' : 'vector';
+      const vecType = columnType.toUpperCase();
       const opclass = useHalfvec ? 'halfvec_cosine_ops' : 'vector_cosine_ops';
+      const hnswMaxDims = hnswMaxDimsForType(columnType);
+      const queryCacheEmbeddingIndexSql = embeddingDim <= hnswMaxDims
+        ? `CREATE INDEX IF NOT EXISTS idx_query_cache_embedding_hnsw
+          ON query_cache USING hnsw (embedding ${opclass})
+          WHERE embedding IS NOT NULL;`
+        : `-- idx_query_cache_embedding_hnsw skipped: pgvector HNSW ${columnType} indexes support
+        -- at most ${hnswMaxDims} dimensions; exact vector scans remain available.`;
 
       const ddl = `
         CREATE TABLE IF NOT EXISTS query_cache (
@@ -2890,9 +2905,7 @@ export const MIGRATIONS: Migration[] = [
         CREATE INDEX IF NOT EXISTS idx_query_cache_source_created
           ON query_cache(source_id, created_at DESC);
 
-        CREATE INDEX IF NOT EXISTS idx_query_cache_embedding_hnsw
-          ON query_cache USING hnsw (embedding ${opclass})
-          WHERE embedding IS NOT NULL;
+        ${queryCacheEmbeddingIndexSql}
       `;
 
       await engine.runMigration(55, ddl);

src/core/pglite-engine.ts

@@ -1058,16 +1058,6 @@ export class PGLiteEngine implements BrainEngine {
        RETURNING id, source_id, slug, type, title, compiled_truth, timeline, frontmatter, content_hash, created_at, updated_at, effective_date, effective_date_source, import_filename, source_kind, source_uri, ingested_via, ingested_at`,
       [sourceId, slug, page.type, pageKind, page.title, page.compiled_truth, page.timeline || '', JSON.stringify(frontmatter), hash, effectiveDate, effectiveDateSource, importFilename, chunkerVersion, sourcePath, sourceKind, sourceUri, ingestedVia, ingestedAt]
     );
-    // #2189: an INSERT … ON CONFLICT DO UPDATE … RETURNING that yields 0 rows
-    // (e.g. a BEFORE trigger suppressing the write) previously crashed in
-    // rowToPage with an opaque "undefined is not an object (row.deleted_at)".
-    // Throw a diagnosable error naming the row instead. Mirrors postgres-engine.ts.
-    if (!rows[0]) {
-      throw new Error(
-        `putPage: INSERT … RETURNING produced no row for slug='${slug}' source_id='${sourceId}'. ` +
-        `A trigger or rule on the pages table may be suppressing the write.`
-      );
-    }
     return rowToPage(rows[0] as Record<string, unknown>);
   }
 

src/core/postgres-engine.ts

@@ -1119,16 +1119,6 @@ export class PostgresEngine implements BrainEngine {
         ingested_at           = COALESCE(EXCLUDED.ingested_at,           pages.ingested_at)
       RETURNING id, source_id, slug, type, title, compiled_truth, timeline, frontmatter, content_hash, created_at, updated_at, effective_date, effective_date_source, import_filename, source_kind, source_uri, ingested_via, ingested_at
     `;
-    // #2189: an INSERT … ON CONFLICT DO UPDATE … RETURNING that yields 0 rows
-    // (e.g. a BEFORE trigger suppressing the write) previously crashed in
-    // rowToPage with an opaque "undefined is not an object (row.deleted_at)".
-    // Throw a diagnosable error naming the row instead. Mirrors pglite-engine.ts.
-    if (!rows[0]) {
-      throw new Error(
-        `putPage: INSERT … RETURNING produced no row for slug='${slug}' source_id='${sourceId}'. ` +
-        `A trigger or rule on the pages table may be suppressing the write.`
-      );
-    }
     return rowToPage(rows[0]);
   }
 

src/core/vector-index.ts

@@ -17,6 +17,7 @@
 import type { BrainEngine } from './engine.ts';
 
 export const PGVECTOR_HNSW_VECTOR_MAX_DIMS = 2000;
+export const PGVECTOR_HNSW_HALFVEC_MAX_DIMS = 4000;
 
 const CHUNK_EMBEDDING_HNSW_INDEX =
   'CREATE INDEX IF NOT EXISTS idx_chunks_embedding ON content_chunks USING hnsw (embedding vector_cosine_ops);';
@@ -29,6 +30,10 @@ export function chunkEmbeddingIndexSql(dims: number): string {
   ].join('\n');
 }
 
+export function hnswMaxDimsForType(columnType: 'vector' | 'halfvec'): number {
+  return columnType === 'halfvec' ? PGVECTOR_HNSW_HALFVEC_MAX_DIMS : PGVECTOR_HNSW_VECTOR_MAX_DIMS;
+}
+
 export function applyChunkEmbeddingIndexPolicy(sql: string, dims: number): string {
   return sql.replaceAll(CHUNK_EMBEDDING_HNSW_INDEX, chunkEmbeddingIndexSql(dims));
 }

test/embedding-dim-check-facts.test.ts

@@ -122,9 +122,9 @@ describe('buildFactsAlterRecipe', () => {
   });
 
   test('vector recipe uses vector_cosine_ops + vector(N) USING cast', () => {
-    const recipe = buildFactsAlterRecipe(1024, 2048, 'vector');
-    expect(recipe).toContain('vector(2048)');
-    expect(recipe).toContain('USING embedding::vector(2048)');
+    const recipe = buildFactsAlterRecipe(1024, 1536, 'vector');
+    expect(recipe).toContain('vector(1536)');
+    expect(recipe).toContain('USING embedding::vector(1536)');
     expect(recipe).toContain('vector_cosine_ops');
     expect(recipe).not.toContain('halfvec_cosine_ops');
   });
@@ -163,6 +163,14 @@ describe('buildFactsAlterRecipe', () => {
     expect(recipe).not.toContain('UPDATE facts SET embedding = NULL');
     expect(recipe).toContain('USING embedding::vector(1536)');
   });
+
+  test('halfvec recipe skips HNSW rebuild above pgvector cap', () => {
+    const recipe = buildFactsAlterRecipe(1536, 4096, 'halfvec');
+    expect(recipe).toContain('halfvec(4096)');
+    expect(recipe).toContain('Skip reindex');
+    expect(recipe).toContain("exceeds pgvector's HNSW cap of 4000");
+    expect(recipe).not.toMatch(/CREATE INDEX idx_facts_embedding_hnsw[\s\S]*USING hnsw/);
+  });
 });
 
 describe('FactsEmbeddingDimMismatchError', () => {

test/facts-migration-dim.test.ts

@@ -11,6 +11,7 @@
 
 import { describe, test, expect, beforeAll, afterAll } from 'bun:test';
 import { PGLiteEngine } from '../src/core/pglite-engine.ts';
+import { configureGateway, resetGateway } from '../src/core/ai/gateway.ts';
 
 let engine: PGLiteEngine;
 
@@ -93,4 +94,60 @@ describe('migration v45 facts column shape', () => {
     );
     expect(after[0].udt_name).toBe(before[0].udt_name);
   });
+
+});
+
+describe('migration v45/v55 large-dim HNSW policy', () => {
+  let largeDimEngine: PGLiteEngine;
+
+  beforeAll(async () => {
+    configureGateway({
+      embedding_model: 'litellm:custom-4096d',
+      embedding_dimensions: 4096,
+      env: { ...process.env },
+    });
+
+    largeDimEngine = new PGLiteEngine();
+    await largeDimEngine.connect({});
+    await largeDimEngine.initSchema();
+  });
+
+  afterAll(async () => {
+    await largeDimEngine.disconnect();
+    resetGateway();
+  });
+
+  test('4096d init skips unsupported HNSW indexes but keeps vector columns', async () => {
+    const formatRows = await largeDimEngine.executeRaw<{ format_type: string }>(
+      `SELECT format_type(atttypid, atttypmod) AS format_type
+       FROM pg_attribute
+       WHERE attrelid = 'facts'::regclass AND attname = 'embedding'`,
+    );
+    expect(formatRows[0]?.format_type).toMatch(/(halfvec|vector)\(4096\)/);
+
+    const indexRows = await largeDimEngine.executeRaw<{ exists: boolean }>(
+      `SELECT EXISTS (
+         SELECT 1 FROM pg_indexes
+         WHERE tablename = 'facts'
+           AND indexname = 'idx_facts_embedding_hnsw'
+       ) AS exists`,
+    );
+    expect(indexRows[0]?.exists).toBe(false);
+
+    const queryCacheFormatRows = await largeDimEngine.executeRaw<{ format_type: string }>(
+      `SELECT format_type(atttypid, atttypmod) AS format_type
+       FROM pg_attribute
+       WHERE attrelid = 'query_cache'::regclass AND attname = 'embedding'`,
+    );
+    expect(queryCacheFormatRows[0]?.format_type).toMatch(/(halfvec|vector)\(4096\)/);
+
+    const queryCacheIndexRows = await largeDimEngine.executeRaw<{ exists: boolean }>(
+      `SELECT EXISTS (
+         SELECT 1 FROM pg_indexes
+         WHERE tablename = 'query_cache'
+           AND indexname = 'idx_query_cache_embedding_hnsw'
+       ) AS exists`,
+    );
+    expect(queryCacheIndexRows[0]?.exists).toBe(false);
+  }, 60000);
 });

test/putpage-returning-guard.test.ts

@@ -1,74 +0,0 @@
-// #2189 regression guard: putPage's INSERT … ON CONFLICT DO UPDATE … RETURNING
-// can yield 0 rows when brain-local DB state (e.g. a BEFORE INSERT trigger)
-// suppresses the write. Pre-fix, rowToPage(rows[0]) crashed with the opaque
-// "undefined is not an object (evaluating 'row.deleted_at')" that failed
-// ~all files of a code sync. Post-fix, putPage throws a descriptive error
-// naming the slug + source_id so the failure is diagnosable per-file.
-//
-// Same guard lands in postgres-engine.ts (engine-parity invariant); this test
-// exercises the PGLite side, where the issue was reported.
-
-import { describe, expect, test, beforeAll, afterAll } from 'bun:test';
-import { PGLiteEngine } from '../src/core/pglite-engine.ts';
-
-let engine: PGLiteEngine;
-
-beforeAll(async () => {
-  engine = new PGLiteEngine();
-  await engine.connect({});
-  await engine.initSchema();
-  // Simulate the reporter's state-dependent failure: a trigger that
-  // suppresses inserts for one slug, making RETURNING produce no row.
-  await engine.executeRaw(`
-    CREATE OR REPLACE FUNCTION suppress_pages_insert() RETURNS trigger AS $$
-    BEGIN
-      IF NEW.slug = 'suppressed-page' THEN RETURN NULL; END IF;
-      RETURN NEW;
-    END;
-    $$ LANGUAGE plpgsql;
-  `);
-  await engine.executeRaw(`
-    CREATE TRIGGER suppress_pages_insert_trg
-    BEFORE INSERT ON pages
-    FOR EACH ROW EXECUTE FUNCTION suppress_pages_insert();
-  `);
-});
-
-afterAll(async () => {
-  await engine.executeRaw('DROP TRIGGER IF EXISTS suppress_pages_insert_trg ON pages');
-  await engine.executeRaw('DROP FUNCTION IF EXISTS suppress_pages_insert');
-  await engine.disconnect();
-});
-
-describe('putPage RETURNING guard (#2189)', () => {
-  test('0-row RETURNING throws a descriptive error, not row.deleted_at TypeError', async () => {
-    let err: Error | undefined;
-    try {
-      await engine.putPage('suppressed-page', {
-        type: 'code',
-        title: 'Suppressed',
-        compiled_truth: 'x',
-        timeline: '',
-      });
-    } catch (e) {
-      err = e as Error;
-    }
-    expect(err).toBeDefined();
-    expect(err!.message).toContain('putPage');
-    expect(err!.message).toContain("slug='suppressed-page'");
-    expect(err!.message).toContain("source_id='default'");
-    // The pre-fix crash signature must be gone.
-    expect(err!.message).not.toContain('deleted_at');
-  });
-
-  test('unsuppressed slugs still upsert normally with the trigger installed', async () => {
-    const page = await engine.putPage('normal-page', {
-      type: 'concept',
-      title: 'Normal',
-      compiled_truth: 'y',
-      timeline: '',
-    });
-    expect(page.slug).toBe('normal-page');
-    expect(page.source_id).toBe('default');
-  });
-});

test/sync.test.ts

@@ -375,31 +375,6 @@ describe('performSync dry-run never writes', () => {
     expect(messages.some(m => m.includes('git pull failed'))).toBe(false);
   });
 
-  test('first PGLite code sync imports code files without runtime failures', async () => {
-    const { performSync } = await import('../src/commands/sync.ts');
-    mkdirSync(join(repoPath, 'src'), { recursive: true });
-    writeFileSync(
-      join(repoPath, 'src/example.ts'),
-      'export function add(left: number, right: number) { return left + right; }\n',
-    );
-    execSync('git add -A && git commit -m "add code file"', { cwd: repoPath, stdio: 'pipe' });
-
-    const result = await performSync(engine, {
-      repoPath,
-      noPull: true,
-      noEmbed: true,
-      noExtract: true,
-      strategy: 'code',
-    });
-
-    expect(result.status).toBe('first_sync');
-    expect(result.added).toBe(1);
-    expect(result.failedFiles ?? 0).toBe(0);
-    const page = await engine.getPage('src-example-ts');
-    expect(page?.type).toBe('code');
-    expect(page?.frontmatter).toMatchObject({ file: 'src/example.ts', language: 'typescript' });
-  });
-
   test('incremental dry-run does NOT write to DB or advance the bookmark', async () => {
     const { performSync } = await import('../src/commands/sync.ts');
     // First do a real sync to seed the bookmark.