diff --git a/sqlite-vec.c b/sqlite-vec.c
index cc09078..a4b9c11 100644
--- a/sqlite-vec.c
+++ b/sqlite-vec.c
@@ -3447,6 +3447,13 @@ static sqlite3_module vec_npy_eachModule = {
 #define VEC0_SHADOW_VECTOR_N_NAME "\"%w\".\"%w_vector_chunks%02d\""
 
 /// 1) schema, 2) original vtab table name
+//
+// IMPORTANT: "rowid" is declared as PRIMARY KEY but WITHOUT the INTEGER type.
+// This means it is NOT a true SQLite rowid alias — the user-defined "rowid"
+// column and the internal SQLite rowid (_rowid_) are two separate values.
+// When inserting, both must be set explicitly to keep them in sync. See the
+// _rowid_ bindings in vec0_new_chunk() and the explanation in
+// SHADOW_TABLE_ROWID_QUIRK below.
 #define VEC0_SHADOW_VECTOR_N_CREATE                                            \
   "CREATE TABLE " VEC0_SHADOW_VECTOR_N_NAME "("                                \
   "rowid PRIMARY KEY,"                                                         \
@@ -4506,6 +4513,20 @@ int vec0_new_chunk(vec0_vtab *p, sqlite3_value ** partitionKeyValues, i64 *chunk
 
   // Step 2: Create new vector chunks for each vector column, with
   //          that new chunk_rowid.
+  //
+  // SHADOW_TABLE_ROWID_QUIRK: The _vector_chunksNN and _metadatachunksNN
+  // shadow tables declare "rowid PRIMARY KEY" without the INTEGER type, so
+  // the user-defined "rowid" column is NOT an alias for the internal SQLite
+  // rowid (_rowid_). When only appending rows these two happen to stay in
+  // sync, but after a chunk is deleted (vec0Update_Delete_DeleteChunkIfEmpty)
+  // and a new one is created, the auto-assigned _rowid_ can diverge from the
+  // user "rowid" value. Since sqlite3_blob_open() addresses rows by internal
+  // _rowid_, we must explicitly set BOTH _rowid_ and "rowid" to the same
+  // value so that later blob operations can find the row.
+  //
+  // The correct long-term fix is changing the schema to
+  //   "rowid INTEGER PRIMARY KEY"
+  // which makes it a true alias, but that would break existing databases.
 
   for (int i = 0; i < vec0_num_defined_user_columns(p); i++) {
     if(p->user_column_kinds[i] != SQLITE_VEC0_USER_COLUMN_KIND_VECTOR) {
@@ -4515,9 +4536,10 @@ int vec0_new_chunk(vec0_vtab *p, sqlite3_value ** partitionKeyValues, i64 *chunk
     i64 vectorsSize =
         p->chunk_size * vector_column_byte_size(p->vector_columns[vector_column_idx]);
 
+    // See SHADOW_TABLE_ROWID_QUIRK above for why _rowid_ and rowid are both set.
     zSql = sqlite3_mprintf("INSERT INTO " VEC0_SHADOW_VECTOR_N_NAME
-                           "(rowid, vectors)"
-                           "VALUES (?, ?)",
+                           "(_rowid_, rowid, vectors)"
+                           "VALUES (?, ?, ?)",
                            p->schemaName, p->tableName, vector_column_idx);
     if (!zSql) {
       return SQLITE_NOMEM;
@@ -4530,8 +4552,9 @@ int vec0_new_chunk(vec0_vtab *p, sqlite3_value ** partitionKeyValues, i64 *chunk
       return rc;
     }
 
-    sqlite3_bind_int64(stmt, 1, rowid);
-    sqlite3_bind_zeroblob64(stmt, 2, vectorsSize);
+    sqlite3_bind_int64(stmt, 1, rowid);  // _rowid_ (internal SQLite rowid)
+    sqlite3_bind_int64(stmt, 2, rowid);  // rowid   (user-defined column)
+    sqlite3_bind_zeroblob64(stmt, 3, vectorsSize);
 
     rc = sqlite3_step(stmt);
     sqlite3_finalize(stmt);
@@ -4546,9 +4569,10 @@ int vec0_new_chunk(vec0_vtab *p, sqlite3_value ** partitionKeyValues, i64 *chunk
       continue;
     }
     int metadata_column_idx = p->user_column_idxs[i];
+    // See SHADOW_TABLE_ROWID_QUIRK above for why _rowid_ and rowid are both set.
     zSql = sqlite3_mprintf("INSERT INTO " VEC0_SHADOW_METADATA_N_NAME
-                           "(rowid, data)"
-                           "VALUES (?, ?)",
+                           "(_rowid_, rowid, data)"
+                           "VALUES (?, ?, ?)",
                            p->schemaName, p->tableName, metadata_column_idx);
     if (!zSql) {
       return SQLITE_NOMEM;
@@ -4561,8 +4585,9 @@ int vec0_new_chunk(vec0_vtab *p, sqlite3_value ** partitionKeyValues, i64 *chunk
       return rc;
     }
 
-    sqlite3_bind_int64(stmt, 1, rowid);
-    sqlite3_bind_zeroblob64(stmt, 2, vec0_metadata_chunk_size(p->metadata_columns[metadata_column_idx].kind, p->chunk_size));
+    sqlite3_bind_int64(stmt, 1, rowid);  // _rowid_ (internal SQLite rowid)
+    sqlite3_bind_int64(stmt, 2, rowid);  // rowid   (user-defined column)
+    sqlite3_bind_zeroblob64(stmt, 3, vec0_metadata_chunk_size(p->metadata_columns[metadata_column_idx].kind, p->chunk_size));
 
     rc = sqlite3_step(stmt);
     sqlite3_finalize(stmt);
@@ -5126,6 +5151,8 @@ static int vec0_init(sqlite3 *db, void *pAux, int argc, const char *const *argv,
       sqlite3_finalize(stmt);
     }
 
+    // See SHADOW_TABLE_ROWID_QUIRK in vec0_new_chunk() — same "rowid PRIMARY KEY"
+    // without INTEGER type issue applies here.
     for (int i = 0; i < pNew->numMetadataColumns; i++) {
       char *zSql = sqlite3_mprintf("CREATE TABLE " VEC0_SHADOW_METADATA_N_NAME "(rowid PRIMARY KEY, data BLOB NOT NULL);",
                                    pNew->schemaName, pNew->tableName, i);
@@ -8574,6 +8601,200 @@ int vec0Update_Delete_ClearValidity(vec0_vtab *p, i64 chunk_id,
   return SQLITE_OK;
 }
 
+int vec0Update_Delete_ClearRowid(vec0_vtab *p, i64 chunk_id,
+                                  u64 chunk_offset) {
+  int rc, brc;
+  sqlite3_blob *blobChunksRowids = NULL;
+  i64 zero = 0;
+
+  rc = sqlite3_blob_open(p->db, p->schemaName, p->shadowChunksName, "rowids",
+                         chunk_id, 1, &blobChunksRowids);
+  if (rc != SQLITE_OK) {
+    vtab_set_error(&p->base, "could not open rowids blob for %s.%s.%lld",
+                   p->schemaName, p->shadowChunksName, chunk_id);
+    return SQLITE_ERROR;
+  }
+
+  rc = sqlite3_blob_write(blobChunksRowids, &zero, sizeof(zero),
+                          chunk_offset * sizeof(i64));
+  if (rc != SQLITE_OK) {
+    vtab_set_error(&p->base,
+                   "could not write to rowids blob for %s.%s.%lld at %llu",
+                   p->schemaName, p->shadowChunksName, chunk_id, chunk_offset);
+  }
+
+  brc = sqlite3_blob_close(blobChunksRowids);
+  if (rc != SQLITE_OK)
+    return rc;
+  if (brc != SQLITE_OK) {
+    vtab_set_error(&p->base,
+                   "vec0 deletion error: Error commiting rowids blob "
+                   "transaction on %s.%s.%lld at %llu",
+                   p->schemaName, p->shadowChunksName, chunk_id, chunk_offset);
+    return brc;
+  }
+  return SQLITE_OK;
+}
+
+int vec0Update_Delete_ClearVectors(vec0_vtab *p, i64 chunk_id,
+                                    u64 chunk_offset) {
+  int rc, brc;
+  for (int i = 0; i < p->numVectorColumns; i++) {
+    sqlite3_blob *blobVectors = NULL;
+    size_t n = vector_column_byte_size(p->vector_columns[i]);
+
+    rc = sqlite3_blob_open(p->db, p->schemaName,
+                           p->shadowVectorChunksNames[i], "vectors",
+                           chunk_id, 1, &blobVectors);
+    if (rc != SQLITE_OK) {
+      vtab_set_error(&p->base,
+                     "could not open vector blob for %s.%s.%lld column %d",
+                     p->schemaName, p->shadowVectorChunksNames[i], chunk_id, i);
+      return SQLITE_ERROR;
+    }
+
+    void *zeroBuf = sqlite3_malloc(n);
+    if (!zeroBuf) {
+      sqlite3_blob_close(blobVectors);
+      return SQLITE_NOMEM;
+    }
+    memset(zeroBuf, 0, n);
+
+    rc = sqlite3_blob_write(blobVectors, zeroBuf, n, chunk_offset * n);
+    sqlite3_free(zeroBuf);
+    if (rc != SQLITE_OK) {
+      vtab_set_error(
+          &p->base,
+          "could not write to vector blob for %s.%s.%lld at %llu column %d",
+          p->schemaName, p->shadowVectorChunksNames[i], chunk_id,
+          chunk_offset, i);
+    }
+
+    brc = sqlite3_blob_close(blobVectors);
+    if (rc != SQLITE_OK)
+      return rc;
+    if (brc != SQLITE_OK) {
+      vtab_set_error(&p->base,
+                     "vec0 deletion error: Error commiting vector blob "
+                     "transaction on %s.%s.%lld column %d",
+                     p->schemaName, p->shadowVectorChunksNames[i], chunk_id, i);
+      return brc;
+    }
+  }
+  return SQLITE_OK;
+}
+
+int vec0Update_Delete_DeleteChunkIfEmpty(vec0_vtab *p, i64 chunk_id,
+                                          int *deleted) {
+  int rc, brc;
+  sqlite3_blob *blobValidity = NULL;
+  *deleted = 0;
+
+  rc = sqlite3_blob_open(p->db, p->schemaName, p->shadowChunksName, "validity",
+                         chunk_id, 0, &blobValidity);
+  if (rc != SQLITE_OK) {
+    vtab_set_error(&p->base,
+                   "could not open validity blob for chunk %lld", chunk_id);
+    return SQLITE_ERROR;
+  }
+
+  int validitySize = sqlite3_blob_bytes(blobValidity);
+  unsigned char *validityBuf = sqlite3_malloc(validitySize);
+  if (!validityBuf) {
+    sqlite3_blob_close(blobValidity);
+    return SQLITE_NOMEM;
+  }
+
+  rc = sqlite3_blob_read(blobValidity, validityBuf, validitySize, 0);
+  brc = sqlite3_blob_close(blobValidity);
+  if (rc != SQLITE_OK) {
+    sqlite3_free(validityBuf);
+    return rc;
+  }
+  if (brc != SQLITE_OK) {
+    sqlite3_free(validityBuf);
+    return brc;
+  }
+
+  int allZero = 1;
+  for (int i = 0; i < validitySize; i++) {
+    if (validityBuf[i] != 0) {
+      allZero = 0;
+      break;
+    }
+  }
+  sqlite3_free(validityBuf);
+
+  if (!allZero) {
+    return SQLITE_OK;
+  }
+
+  // All validity bits are zero — delete this chunk and its associated data
+  char *zSql;
+  sqlite3_stmt *stmt;
+
+  // Delete from _chunks
+  zSql = sqlite3_mprintf(
+      "DELETE FROM " VEC0_SHADOW_CHUNKS_NAME " WHERE rowid = ?",
+      p->schemaName, p->tableName);
+  if (!zSql)
+    return SQLITE_NOMEM;
+  rc = sqlite3_prepare_v2(p->db, zSql, -1, &stmt, NULL);
+  sqlite3_free(zSql);
+  if (rc != SQLITE_OK)
+    return rc;
+  sqlite3_bind_int64(stmt, 1, chunk_id);
+  rc = sqlite3_step(stmt);
+  sqlite3_finalize(stmt);
+  if (rc != SQLITE_DONE)
+    return SQLITE_ERROR;
+
+  // Delete from each _vector_chunksNN
+  for (int i = 0; i < p->numVectorColumns; i++) {
+    zSql = sqlite3_mprintf(
+        "DELETE FROM " VEC0_SHADOW_VECTOR_N_NAME " WHERE rowid = ?",
+        p->schemaName, p->tableName, i);
+    if (!zSql)
+      return SQLITE_NOMEM;
+    rc = sqlite3_prepare_v2(p->db, zSql, -1, &stmt, NULL);
+    sqlite3_free(zSql);
+    if (rc != SQLITE_OK)
+      return rc;
+    sqlite3_bind_int64(stmt, 1, chunk_id);
+    rc = sqlite3_step(stmt);
+    sqlite3_finalize(stmt);
+    if (rc != SQLITE_DONE)
+      return SQLITE_ERROR;
+  }
+
+  // Delete from each _metadatachunksNN
+  for (int i = 0; i < p->numMetadataColumns; i++) {
+    zSql = sqlite3_mprintf(
+        "DELETE FROM " VEC0_SHADOW_METADATA_N_NAME " WHERE rowid = ?",
+        p->schemaName, p->tableName, i);
+    if (!zSql)
+      return SQLITE_NOMEM;
+    rc = sqlite3_prepare_v2(p->db, zSql, -1, &stmt, NULL);
+    sqlite3_free(zSql);
+    if (rc != SQLITE_OK)
+      return rc;
+    sqlite3_bind_int64(stmt, 1, chunk_id);
+    rc = sqlite3_step(stmt);
+    sqlite3_finalize(stmt);
+    if (rc != SQLITE_DONE)
+      return SQLITE_ERROR;
+  }
+
+  // Invalidate cached stmtLatestChunk so it gets re-prepared on next insert
+  if (p->stmtLatestChunk) {
+    sqlite3_finalize(p->stmtLatestChunk);
+    p->stmtLatestChunk = NULL;
+  }
+
+  *deleted = 1;
+  return SQLITE_OK;
+}
+
 int vec0Update_Delete_DeleteRowids(vec0_vtab *p, i64 rowid) {
   int rc;
   sqlite3_stmt *stmt = NULL;
@@ -8735,16 +8956,23 @@ int vec0Update_Delete(sqlite3_vtab *pVTab, sqlite3_value *idValue) {
     return rc;
   }
 
+  // 2. clear validity bit
   rc = vec0Update_Delete_ClearValidity(p, chunk_id, chunk_offset);
   if (rc != SQLITE_OK) {
     return rc;
   }
 
   // 3. zero out rowid in chunks.rowids
-  // https://github.com/asg017/sqlite-vec/issues/54
+  rc = vec0Update_Delete_ClearRowid(p, chunk_id, chunk_offset);
+  if (rc != SQLITE_OK) {
+    return rc;
+  }
 
   // 4. zero out any data in vector chunks tables
-  // https://github.com/asg017/sqlite-vec/issues/54
+  rc = vec0Update_Delete_ClearVectors(p, chunk_id, chunk_offset);
+  if (rc != SQLITE_OK) {
+    return rc;
+  }
 
   // 5. delete from _rowids table
   rc = vec0Update_Delete_DeleteRowids(p, rowid);
@@ -8760,9 +8988,21 @@ int vec0Update_Delete(sqlite3_vtab *pVTab, sqlite3_value *idValue) {
     }
   }
 
-  // 6. delete metadata
+  // 7. delete metadata
   for(int i = 0; i < p->numMetadataColumns; i++) {
     rc = vec0Update_Delete_ClearMetadata(p, i, rowid, chunk_id, chunk_offset);
+    if (rc != SQLITE_OK) {
+      return rc;
+    }
+  }
+
+  // 8. reclaim chunk if fully empty
+  {
+    int chunkDeleted;
+    rc = vec0Update_Delete_DeleteChunkIfEmpty(p, chunk_id, &chunkDeleted);
+    if (rc != SQLITE_OK) {
+      return rc;
+    }
   }
 
   return SQLITE_OK;
diff --git a/tests/__snapshots__/test-auxiliary.ambr b/tests/__snapshots__/test-auxiliary.ambr
index bfe3d2c..66a3ef3 100644
--- a/tests/__snapshots__/test-auxiliary.ambr
+++ b/tests/__snapshots__/test-auxiliary.ambr
@@ -137,7 +137,7 @@
           'chunk_id': 1,
           'size': 8,
           'validity': b'\x06',
-          'rowids': b'\x01\x00\x00\x00\x00\x00\x00\x00\x02\x00\x00\x00\x00\x00\x00\x00\x03\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00',
+          'rowids': b'\x00\x00\x00\x00\x00\x00\x00\x00\x02\x00\x00\x00\x00\x00\x00\x00\x03\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00',
         }),
       ]),
     }),
@@ -163,7 +163,7 @@
       'rows': list([
         OrderedDict({
           'rowid': 1,
-          'vectors': b'\x00\x00\x80?\x00\x00\x00@\x00\x00@@\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00',
+          'vectors': b'\x00\x00\x00\x00\x00\x00\x00@\x00\x00@@\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00',
         }),
       ]),
     }),
diff --git a/tests/__snapshots__/test-metadata.ambr b/tests/__snapshots__/test-metadata.ambr
index 12212ff..ff7b112 100644
--- a/tests/__snapshots__/test-metadata.ambr
+++ b/tests/__snapshots__/test-metadata.ambr
@@ -27,8 +27,8 @@
         OrderedDict({
           'chunk_id': 1,
           'size': 8,
-          'validity': b'\x02',
-          'rowids': b'\x01\x00\x00\x00\x00\x00\x00\x00\x02\x00\x00\x00\x00\x00\x00\x00\x03\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00',
+          'validity': b'\x06',
+          'rowids': b'\x00\x00\x00\x00\x00\x00\x00\x00\x02\x00\x00\x00\x00\x00\x00\x00\x03\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00',
         }),
       ]),
     }),
@@ -37,7 +37,7 @@
       'rows': list([
         OrderedDict({
           'rowid': 1,
-          'data': b'\x02',
+          'data': b'\x06',
         }),
       ]),
     }),
@@ -46,7 +46,7 @@
       'rows': list([
         OrderedDict({
           'rowid': 1,
-          'data': b'\x00\x00\x00\x00\x00\x00\x00\x00\x02\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00',
+          'data': b'\x00\x00\x00\x00\x00\x00\x00\x00\x02\x00\x00\x00\x00\x00\x00\x00\x03\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00',
         }),
       ]),
     }),
@@ -55,7 +55,7 @@
       'rows': list([
         OrderedDict({
           'rowid': 1,
-          'data': b'\x00\x00\x00\x00\x00\x00\x00\x00\x9a\x99\x99\x99\x99\x99\x01@\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00',
+          'data': b'\x00\x00\x00\x00\x00\x00\x00\x00\x9a\x99\x99\x99\x99\x99\x01@ffffff\n@\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00',
         }),
       ]),
     }),
@@ -64,13 +64,17 @@
       'rows': list([
         OrderedDict({
           'rowid': 1,
-          'data': b'\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x05\x00\x00\x00test2\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00',
+          'data': b'\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x05\x00\x00\x00test2\x00\x00\x00\x00\x00\x00\x00\r\x00\x00\x00123456789012\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00',
         }),
       ]),
     }),
     'v_metadatatext03': OrderedDict({
       'sql': 'select * from v_metadatatext03',
       'rows': list([
+        OrderedDict({
+          'rowid': 3,
+          'data': '1234567890123',
+        }),
       ]),
     }),
     'v_rowids': OrderedDict({
@@ -82,6 +86,12 @@
           'chunk_id': 1,
           'chunk_offset': 1,
         }),
+        OrderedDict({
+          'rowid': 3,
+          'id': None,
+          'chunk_id': 1,
+          'chunk_offset': 2,
+        }),
       ]),
     }),
     'v_vector_chunks00': OrderedDict({
@@ -89,7 +99,7 @@
       'rows': list([
         OrderedDict({
           'rowid': 1,
-          'vectors': b'\x11\x11\x11\x11""""3333\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00',
+          'vectors': b'\x00\x00\x00\x00""""3333\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00',
         }),
       ]),
     }),
@@ -264,7 +274,7 @@
           'chunk_id': 1,
           'size': 8,
           'validity': b'\x06',
-          'rowids': b'\x01\x00\x00\x00\x00\x00\x00\x00\x02\x00\x00\x00\x00\x00\x00\x00\x03\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00',
+          'rowids': b'\x00\x00\x00\x00\x00\x00\x00\x00\x02\x00\x00\x00\x00\x00\x00\x00\x03\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00',
         }),
       ]),
     }),
@@ -335,7 +345,7 @@
       'rows': list([
         OrderedDict({
           'rowid': 1,
-          'vectors': b'\x11\x11\x11\x11""""3333\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00',
+          'vectors': b'\x00\x00\x00\x00""""3333\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00',
         }),
       ]),
     }),
@@ -360,6 +370,14 @@
         'f': 2.2,
         't': 'test2',
       }),
+      OrderedDict({
+        'rowid': 3,
+        'vector': b'3333',
+        'b': 1,
+        'n': 3,
+        'f': 3.3,
+        't': '1234567890123',
+      }),
     ]),
   })
 # ---
diff --git a/tests/fuzz/Makefile b/tests/fuzz/Makefile
index 0f1e5ba..21629ef 100644
--- a/tests/fuzz/Makefile
+++ b/tests/fuzz/Makefile
@@ -69,9 +69,13 @@ $(TARGET_DIR)/vec_each: vec-each.c $(FUZZ_SRCS) | $(TARGET_DIR)
 $(TARGET_DIR)/vec_mismatch: vec-mismatch.c $(FUZZ_SRCS) | $(TARGET_DIR)
 	$(FUZZ_CC) $(FUZZ_CFLAGS) $(FUZZ_SRCS) $< -o $@
 
+$(TARGET_DIR)/vec0_delete_completeness: vec0-delete-completeness.c $(FUZZ_SRCS) | $(TARGET_DIR)
+	$(FUZZ_CC) $(FUZZ_CFLAGS) $(FUZZ_SRCS) $< -o $@
+
 FUZZ_TARGETS = vec0_create exec json numpy \
                shadow_corrupt vec0_operations scalar_functions \
-               vec0_create_full metadata_columns vec_each vec_mismatch
+               vec0_create_full metadata_columns vec_each vec_mismatch \
+               vec0_delete_completeness
 
 all: $(addprefix $(TARGET_DIR)/,$(FUZZ_TARGETS))
 
diff --git a/tests/fuzz/vec0-delete-completeness.c b/tests/fuzz/vec0-delete-completeness.c
new file mode 100644
index 0000000..8603b71
--- /dev/null
+++ b/tests/fuzz/vec0-delete-completeness.c
@@ -0,0 +1,114 @@
+#include <stdint.h>
+#include <stddef.h>
+#include <stdio.h>
+#include <stdlib.h>
+#include <string.h>
+#include "sqlite-vec.h"
+#include "sqlite3.h"
+#include <assert.h>
+
+int LLVMFuzzerTestOneInput(const uint8_t *data, size_t size) {
+  if (size < 6) return 0;
+
+  int rc;
+  sqlite3 *db;
+  sqlite3_stmt *stmtInsert = NULL;
+  sqlite3_stmt *stmtDelete = NULL;
+  sqlite3_stmt *stmtScan = NULL;
+  sqlite3_stmt *stmtCount = NULL;
+
+  rc = sqlite3_open(":memory:", &db);
+  assert(rc == SQLITE_OK);
+  rc = sqlite3_vec_init(db, NULL, NULL);
+  assert(rc == SQLITE_OK);
+
+  rc = sqlite3_exec(db,
+    "CREATE VIRTUAL TABLE v USING vec0(emb float[4], chunk_size=4)",
+    NULL, NULL, NULL);
+  if (rc != SQLITE_OK) { sqlite3_close(db); return 0; }
+
+  sqlite3_prepare_v2(db,
+    "INSERT INTO v(rowid, emb) VALUES (?, ?)", -1, &stmtInsert, NULL);
+  sqlite3_prepare_v2(db,
+    "DELETE FROM v WHERE rowid = ?", -1, &stmtDelete, NULL);
+  sqlite3_prepare_v2(db,
+    "SELECT rowid FROM v", -1, &stmtScan, NULL);
+
+  if (!stmtInsert || !stmtDelete || !stmtScan) goto cleanup;
+
+  size_t i = 0;
+  while (i + 2 <= size) {
+    uint8_t op = data[i++] % 3;
+    uint8_t rowid_byte = data[i++];
+    int64_t rowid = (int64_t)(rowid_byte % 16) + 1;
+
+    switch (op) {
+      case 0: {
+        // INSERT
+        float vec[4] = {0.0f, 0.0f, 0.0f, 0.0f};
+        for (int j = 0; j < 4 && i < size; j++, i++) {
+          vec[j] = (float)((int8_t)data[i]) / 10.0f;
+        }
+        sqlite3_reset(stmtInsert);
+        sqlite3_bind_int64(stmtInsert, 1, rowid);
+        sqlite3_bind_blob(stmtInsert, 2, vec, sizeof(vec), SQLITE_TRANSIENT);
+        sqlite3_step(stmtInsert);
+        break;
+      }
+      case 1: {
+        // DELETE
+        sqlite3_reset(stmtDelete);
+        sqlite3_bind_int64(stmtDelete, 1, rowid);
+        sqlite3_step(stmtDelete);
+        break;
+      }
+      case 2: {
+        // Full scan
+        sqlite3_reset(stmtScan);
+        while (sqlite3_step(stmtScan) == SQLITE_ROW) {}
+        break;
+      }
+    }
+  }
+
+  // Delete all remaining rows
+  sqlite3_exec(db, "DELETE FROM v", NULL, NULL, NULL);
+
+  // Assert all shadow tables are empty after full deletion
+  sqlite3_prepare_v2(db,
+    "SELECT count(*) FROM v_rowids", -1, &stmtCount, NULL);
+  if (stmtCount) {
+    rc = sqlite3_step(stmtCount);
+    assert(rc == SQLITE_ROW);
+    assert(sqlite3_column_int(stmtCount, 0) == 0);
+    sqlite3_finalize(stmtCount);
+    stmtCount = NULL;
+  }
+
+  sqlite3_prepare_v2(db,
+    "SELECT count(*) FROM v_chunks", -1, &stmtCount, NULL);
+  if (stmtCount) {
+    rc = sqlite3_step(stmtCount);
+    assert(rc == SQLITE_ROW);
+    assert(sqlite3_column_int(stmtCount, 0) == 0);
+    sqlite3_finalize(stmtCount);
+    stmtCount = NULL;
+  }
+
+  sqlite3_prepare_v2(db,
+    "SELECT count(*) FROM v_vector_chunks00", -1, &stmtCount, NULL);
+  if (stmtCount) {
+    rc = sqlite3_step(stmtCount);
+    assert(rc == SQLITE_ROW);
+    assert(sqlite3_column_int(stmtCount, 0) == 0);
+    sqlite3_finalize(stmtCount);
+    stmtCount = NULL;
+  }
+
+cleanup:
+  sqlite3_finalize(stmtInsert);
+  sqlite3_finalize(stmtDelete);
+  sqlite3_finalize(stmtScan);
+  sqlite3_close(db);
+  return 0;
+}
diff --git a/tests/test-insert-delete.py b/tests/test-insert-delete.py
index 343c231..cc1697b 100644
--- a/tests/test-insert-delete.py
+++ b/tests/test-insert-delete.py
@@ -1,7 +1,7 @@
 import sqlite3
 import struct
 import pytest
-from helpers import _f32, exec
+from helpers import _f32, _i64, _int8, exec
 
 
 def test_insert_creates_chunks_and_vectors(db, snapshot):
@@ -147,3 +147,335 @@ def test_insert_validates_type(db):
 def test_info_table_contents(db, snapshot):
     db.execute("create virtual table v using vec0(emb float[4], chunk_size=8)")
     assert exec(db, "select key, value from v_info order by key") == snapshot()
+
+
+def test_delete_zeroes_rowid_blob(db):
+    db.execute("create virtual table v using vec0(emb float[4], chunk_size=8)")
+
+    for i in range(1, 4):
+        db.execute(
+            "insert into v(rowid, emb) values (?, ?)",
+            [i, _f32([float(i)] * 4)],
+        )
+
+    db.execute("delete from v where rowid = 2")
+
+    blob = db.execute("select rowids from v_chunks where rowid = 1").fetchone()[0]
+    rowids = struct.unpack("<8q", blob)
+    assert rowids[0] == 1  # slot 0 intact
+    assert rowids[1] == 0  # slot 1 zeroed (was rowid 2)
+    assert rowids[2] == 3  # slot 2 intact
+
+
+def test_delete_zeroes_vector_blob(db):
+    db.execute("create virtual table v using vec0(emb float[4], chunk_size=8)")
+
+    db.execute(
+        "insert into v(rowid, emb) values (1, ?)", [_f32([1.0, 2.0, 3.0, 4.0])]
+    )
+    db.execute(
+        "insert into v(rowid, emb) values (2, ?)", [_f32([5.0, 6.0, 7.0, 8.0])]
+    )
+
+    db.execute("delete from v where rowid = 1")
+
+    blob = db.execute(
+        "select vectors from v_vector_chunks00 where rowid = 1"
+    ).fetchone()[0]
+    # First slot (4 floats = 16 bytes) should be zeroed
+    first_slot = struct.unpack("<4f", blob[:16])
+    assert first_slot == (0.0, 0.0, 0.0, 0.0)
+    # Second slot should be unchanged
+    second_slot = struct.unpack("<4f", blob[16:32])
+    assert second_slot == (5.0, 6.0, 7.0, 8.0)
+
+
+def test_delete_all_rows_deletes_chunk(db):
+    db.execute("create virtual table v using vec0(emb float[4], chunk_size=8)")
+
+    for i in range(1, 9):
+        db.execute(
+            "insert into v(rowid, emb) values (?, ?)",
+            [i, _f32([float(i)] * 4)],
+        )
+
+    for i in range(1, 9):
+        db.execute("delete from v where rowid = ?", [i])
+
+    assert (
+        db.execute("select count(*) from v_chunks").fetchone()[0] == 0
+    )
+    assert (
+        db.execute("select count(*) from v_vector_chunks00").fetchone()[0] == 0
+    )
+
+    # Inserting after full deletion still works
+    db.execute(
+        "insert into v(rowid, emb) values (100, ?)", [_f32([9.0, 9.0, 9.0, 9.0])]
+    )
+    row = db.execute("select emb from v where rowid = 100").fetchone()
+    assert row[0] == _f32([9.0, 9.0, 9.0, 9.0])
+
+
+def test_delete_chunk_multiple_chunks(db):
+    db.execute("create virtual table v using vec0(emb float[4], chunk_size=8)")
+
+    for i in range(1, 17):
+        db.execute(
+            "insert into v(rowid, emb) values (?, ?)",
+            [i, _f32([float(i)] * 4)],
+        )
+
+    # Delete all rows from the first chunk (rows 1-8)
+    for i in range(1, 9):
+        db.execute("delete from v where rowid = ?", [i])
+
+    # Only 1 chunk should remain
+    assert db.execute("select count(*) from v_chunks").fetchone()[0] == 1
+
+    # Rows 9-16 still queryable
+    for i in range(9, 17):
+        row = db.execute("select emb from v where rowid = ?", [i]).fetchone()
+        assert row[0] == _f32([float(i)] * 4)
+
+
+def test_delete_with_metadata_columns(db):
+    db.execute(
+        "create virtual table v using vec0("
+        "emb float[4], "
+        "m_bool boolean, "
+        "m_int integer, "
+        "m_float float, "
+        "m_text text, "
+        "chunk_size=8"
+        ")"
+    )
+
+    for i in range(1, 9):
+        db.execute(
+            "insert into v(rowid, emb, m_bool, m_int, m_float, m_text) "
+            "values (?, ?, ?, ?, ?, ?)",
+            [i, _f32([float(i)] * 4), i % 2 == 0, i * 10, float(i) / 2.0, f"text_{i}"],
+        )
+
+    for i in range(1, 9):
+        db.execute("delete from v where rowid = ?", [i])
+
+    assert db.execute("select count(*) from v_chunks").fetchone()[0] == 0
+    assert db.execute("select count(*) from v_vector_chunks00").fetchone()[0] == 0
+    assert db.execute("select count(*) from v_metadatachunks00").fetchone()[0] == 0
+    assert db.execute("select count(*) from v_metadatachunks01").fetchone()[0] == 0
+    assert db.execute("select count(*) from v_metadatachunks02").fetchone()[0] == 0
+    assert db.execute("select count(*) from v_metadatachunks03").fetchone()[0] == 0
+
+
+def test_delete_with_auxiliary_columns(db):
+    db.execute(
+        "create virtual table v using vec0("
+        "emb float[4], "
+        "+aux_text text, "
+        "chunk_size=8"
+        ")"
+    )
+
+    for i in range(1, 9):
+        db.execute(
+            "insert into v(rowid, emb, aux_text) values (?, ?, ?)",
+            [i, _f32([float(i)] * 4), f"aux_{i}"],
+        )
+
+    for i in range(1, 9):
+        db.execute("delete from v where rowid = ?", [i])
+
+    assert db.execute("select count(*) from v_chunks").fetchone()[0] == 0
+    assert db.execute("select count(*) from v_auxiliary").fetchone()[0] == 0
+
+
+def test_delete_with_text_primary_key(db):
+    db.execute(
+        "create virtual table v using vec0("
+        "id text primary key, emb float[4], chunk_size=8"
+        ")"
+    )
+
+    db.execute(
+        "insert into v(id, emb) values ('a', ?)", [_f32([1.0, 2.0, 3.0, 4.0])]
+    )
+    db.execute(
+        "insert into v(id, emb) values ('b', ?)", [_f32([5.0, 6.0, 7.0, 8.0])]
+    )
+
+    db.execute("delete from v where id = 'a'")
+
+    # Vector blob slot 0 should be zeroed
+    blob = db.execute(
+        "select vectors from v_vector_chunks00 where rowid = 1"
+    ).fetchone()[0]
+    first_slot = struct.unpack("<4f", blob[:16])
+    assert first_slot == (0.0, 0.0, 0.0, 0.0)
+
+    # Remaining row still queryable
+    row = db.execute("select emb from v where id = 'b'").fetchone()
+    assert row[0] == _f32([5.0, 6.0, 7.0, 8.0])
+
+
+def test_delete_with_partition_keys(db):
+    db.execute(
+        "create virtual table v using vec0("
+        "part text partition key, emb float[4], chunk_size=8"
+        ")"
+    )
+
+    for i in range(1, 9):
+        db.execute(
+            "insert into v(rowid, part, emb) values (?, 'A', ?)",
+            [i, _f32([float(i)] * 4)],
+        )
+    for i in range(9, 17):
+        db.execute(
+            "insert into v(rowid, part, emb) values (?, 'B', ?)",
+            [i, _f32([float(i)] * 4)],
+        )
+
+    # Delete all from partition A
+    for i in range(1, 9):
+        db.execute("delete from v where rowid = ?", [i])
+
+    # 1 chunk should remain (partition B's)
+    assert db.execute("select count(*) from v_chunks").fetchone()[0] == 1
+
+    # Partition B rows intact
+    for i in range(9, 17):
+        row = db.execute("select emb from v where rowid = ?", [i]).fetchone()
+        assert row[0] == _f32([float(i)] * 4)
+
+    # Re-insert into partition A works
+    db.execute(
+        "insert into v(rowid, part, emb) values (100, 'A', ?)",
+        [_f32([99.0, 99.0, 99.0, 99.0])],
+    )
+    row = db.execute("select emb from v where rowid = 100").fetchone()
+    assert row[0] == _f32([99.0, 99.0, 99.0, 99.0])
+
+
+def test_delete_int8_vectors(db):
+    db.execute("create virtual table v using vec0(emb int8[4], chunk_size=8)")
+
+    db.execute(
+        "insert into v(rowid, emb) values (1, vec_int8(?))",
+        [_int8([1, 2, 3, 4])],
+    )
+    db.execute(
+        "insert into v(rowid, emb) values (2, vec_int8(?))",
+        [_int8([5, 6, 7, 8])],
+    )
+
+    db.execute("delete from v where rowid = 1")
+
+    blob = db.execute(
+        "select vectors from v_vector_chunks00 where rowid = 1"
+    ).fetchone()[0]
+    # int8[4] = 4 bytes per slot
+    first_slot = struct.unpack("<4b", blob[:4])
+    assert first_slot == (0, 0, 0, 0)
+    second_slot = struct.unpack("<4b", blob[4:8])
+    assert second_slot == (5, 6, 7, 8)
+
+
+def test_delete_bit_vectors(db):
+    db.execute("create virtual table v using vec0(emb bit[8], chunk_size=8)")
+
+    db.execute(
+        "insert into v(rowid, emb) values (1, vec_bit(?))",
+        [bytes([0xFF])],
+    )
+    db.execute(
+        "insert into v(rowid, emb) values (2, vec_bit(?))",
+        [bytes([0xAA])],
+    )
+
+    db.execute("delete from v where rowid = 1")
+
+    blob = db.execute(
+        "select vectors from v_vector_chunks00 where rowid = 1"
+    ).fetchone()[0]
+    # bit[8] = 1 byte per slot
+    assert blob[0:1] == bytes([0x00])
+    assert blob[1:2] == bytes([0xAA])
+
+
+def _file_db(tmp_path):
+    """Open a file-backed DB (required for page_count to shrink after VACUUM)."""
+    db = sqlite3.connect(str(tmp_path / "test.db"))
+    db.row_factory = sqlite3.Row
+    db.enable_load_extension(True)
+    db.load_extension("dist/vec0")
+    db.enable_load_extension(False)
+    return db
+
+
+def test_delete_chunk_shrinks_pages(tmp_path):
+    """Use large vectors (float[256]) so each chunk blob spans multiple pages,
+    making the page_count difference measurable after VACUUM."""
+    dims = 256
+    db = _file_db(tmp_path)
+    db.execute(f"create virtual table v using vec0(emb float[{dims}], chunk_size=8)")
+
+    for i in range(1, 25):  # 3 full chunks of 8
+        db.execute(
+            "insert into v(rowid, emb) values (?, ?)",
+            [i, _f32([float(i)] * dims)],
+        )
+    db.commit()
+    pages_before = db.execute("pragma page_count").fetchone()[0]
+
+    # Delete all rows
+    for i in range(1, 25):
+        db.execute("delete from v where rowid = ?", [i])
+    db.commit()
+
+    assert db.execute("select count(*) from v_chunks").fetchone()[0] == 0
+
+    db.execute("vacuum")
+    pages_after = db.execute("pragma page_count").fetchone()[0]
+    assert pages_after < pages_before, (
+        f"page_count should shrink after deleting all chunks and vacuum: "
+        f"{pages_before} -> {pages_after}"
+    )
+    db.close()
+
+
+def test_delete_one_chunk_of_two_shrinks_pages(tmp_path):
+    """Use large vectors (float[256]) so each chunk blob spans multiple pages,
+    making the page_count difference measurable after VACUUM."""
+    dims = 256
+    db = _file_db(tmp_path)
+    db.execute(f"create virtual table v using vec0(emb float[{dims}], chunk_size=8)")
+
+    for i in range(1, 17):  # 2 full chunks of 8
+        db.execute(
+            "insert into v(rowid, emb) values (?, ?)",
+            [i, _f32([float(i)] * dims)],
+        )
+    db.commit()
+    pages_before = db.execute("pragma page_count").fetchone()[0]
+
+    # Delete all rows from the first chunk (rows 1-8)
+    for i in range(1, 9):
+        db.execute("delete from v where rowid = ?", [i])
+    db.commit()
+
+    assert db.execute("select count(*) from v_chunks").fetchone()[0] == 1
+
+    db.execute("vacuum")
+    pages_after = db.execute("pragma page_count").fetchone()[0]
+    assert pages_after < pages_before, (
+        f"page_count should shrink after deleting one chunk and vacuum: "
+        f"{pages_before} -> {pages_after}"
+    )
+
+    # Remaining rows still queryable after vacuum
+    for i in range(9, 17):
+        row = db.execute("select emb from v where rowid = ?", [i]).fetchone()
+        assert row[0] == _f32([float(i)] * dims)
+    db.close()