feat: support pyramid raw vector retrieval

docs/docs/en/src/advanced/new_serialization.md

@@ -207,6 +207,7 @@ Pyramid writes these streaming blocks in order:
 | `label_table` | external labels and label remap | yes |
 | `base_codes` | base flatten codes used by graph search | yes |
 | `high_precision_codes` | precise reorder codes when reorder is enabled | conditional |
+| `raw_vector` | separate FP32 vectors when `store_raw_vector` requires dedicated storage | conditional |
 | `pyramid_hierarchies` | hierarchy names and graph roots | yes |
 
 `DeserializeStreaming` restores the full in-memory Pyramid index. `Index::Load` can create the

docs/docs/en/src/indexes/pyramid.md

@@ -98,6 +98,9 @@ Build-time parameters live under `index_param`.
 | `fast_encode_rabitq_rounds` | int | `6` | Fast RaBitQ refinement rounds in `[1, 32]`. |
 | `base_io_type` / `precise_io_type` | string | `"block_memory_io"` | Base and reorder storage backends; `uring_io` is available in builds with liburing. |
 | `base_file_path` / `precise_file_path` | string | — | Required for disk-backed storage such as `buffer_io`, `async_io`, `uring_io`, or `mmap_io`. |
+| `store_raw_vector` | bool | `false` | Preserve an FP32 copy for `GetRawVectorByIds` and precise distance-by-id calculations. Pyramid reuses an existing in-memory FP32 base/reorder copy when possible; otherwise it creates separate raw-vector storage. |
+| `raw_vector_io_type` | string | `"block_memory_io"` | IO backend for separate raw-vector storage. |
+| `raw_vector_file_path` | string | `"./default_file_path"` | File path used by file-backed raw-vector IO. |
 | `index_min_size` | int | `0` | Minimum sub-index size; smaller groups fall back to scan. |
 | `support_duplicate` | bool | `false` | Allow duplicate ids. |
 | `build_thread_count` | int | `1` | Threads used for parallel build. |

docs/docs/zh/src/advanced/new_serialization.md

@@ -187,6 +187,7 @@ Pyramid 按顺序写入以下 streaming blocks：
 | `label_table` | 外部 label 和 label remap | 是 |
 | `base_codes` | 图搜索使用的 base flatten codes | 是 |
 | `high_precision_codes` | reorder 开启时的精排 codes | 条件必需 |
+| `raw_vector` | `store_raw_vector` 需要独立存储时的 FP32 向量 | 条件必需 |
 | `pyramid_hierarchies` | hierarchy 名称和 graph roots | 是 |
 
 `DeserializeStreaming` 会恢复完整的内存 Pyramid 索引。`Index::Load` 可以直接从 streaming metadata

docs/docs/zh/src/indexes/pyramid.md

@@ -93,6 +93,9 @@ auto result = index->KnnSearch(
 | `fast_encode_rabitq_rounds` | int | `6` | RaBitQ 快速编码的微调轮数，范围 `[1, 32]` |
 | `base_io_type` / `precise_io_type` | string | `"block_memory_io"` | 底层与精排存储后端；以 liburing 构建时可用 `uring_io` |
 | `base_file_path` / `precise_file_path` | string | — | `buffer_io`、`async_io`、`uring_io`、`mmap_io` 等磁盘存储必须设置 |
+| `store_raw_vector` | bool | `false` | 保留 FP32 向量，供 `GetRawVectorByIds` 和精确的按 ID 距离计算使用。若已有内存 FP32 base/reorder 副本，Pyramid 会直接复用；否则创建独立 raw-vector 存储。 |
+| `raw_vector_io_type` | string | `"block_memory_io"` | 独立 raw-vector 存储使用的 IO 后端 |
+| `raw_vector_file_path` | string | `"./default_file_path"` | 文件型 raw-vector IO 使用的文件路径 |
 | `index_min_size` | int | `0` | 子索引的最小规模；小于该值的分区会退化为线性扫描 |
 | `support_duplicate` | bool | `false` | 是否允许重复 ID |
 | `build_thread_count` | int | `1` | 构建阶段并发线程数 |