Item datasets are the core data foundation driving the platform's precise search, recommendation, and conversational capabilities. They are ideal for scenarios such as building intelligent search for video portals or retrieval systems for media libraries. This chapter provides detailed operational guidance to help you import your video data step\-by\-step and build a structured video dataset, covering key stages such as data uploading, information configuration, and strategy setting. The preparation and creation method for a dataset depends on the data type you are importing. This document focuses on the **Video Data** creation workflow, including data format requirements, field mapping rules, and solutions to common issues. If you wish to create an Image & Text dataset, please refer to ** ** [Create Item Dataset - Image & Text Data](/docs/85296/1544952). :::tip \*Note This chapter primarily focuses on operational workflows for dataset creation. If you have questions regarding theoretical content — such as basic concepts of item datasets, supported data formats, specific definitions of field types, detailed parameter explanations, or search strategies — please first refer to [Dataset Introduction](/docs/85296/1544957) chapter. Understanding the fundamentals will help you complete this chapter's practical operations more smoothly and effectively. ::: # Video Dataset ## Data Preparation Before you begin uploading a video dataset, it is essential to prepare your media data according to the following requirements. Unlike Image/Text datasets, video datasets include several **key fields that must be uploaded using fixed field names**. Therefore, it is critical to verify and adjust your data before the upload process. Please note the following restrictions before uploading a video dataset: :::tip \*说明 * **Two\-Layer Hierarchy Only:** Video content levels only support a 2\-layer structure; a parent item cannot be a child of another item. If a media item already serves as a parent containing children, providing a `parent_content_id` for that parent will result in a processing error. For a detailed introduction, see [Video Data Hierarchy](/docs/85296/1544957#e5817a8e). * **Sequence Ordering:** You can use `sequence_index` to define the order of child content within a parent. If the `sequence_index` is disordered (e.g., duplicate indices or missing values), the system will ignore the ordering data, which may disrupt the chronological flow and affect video understanding. * **4\-Hour Limit:** Single videos exceeding **4 hours** are not supported. ::: **Field name**: Must start with a letter and can only contain letters, numbers, and underscores. **Field type**: Supports JSON schema types: * String (Array) * Integer (Array) — supports Int64 and Int32 * Float (Array) * Boolean * Object (Array) — supports nested objects up to three levels of hierarchy **Field meaning**: During the intelligent schema parsing phase, the LLM will interpret the meaning and content of each field based on the field name and uploaded sample data. These meanings, along with the field values, serve as semantic information within the *vector representation*. This enhances the accuracy of supplementary context for field content during semantic searches. We recommend providing a concise description of the field's meaning in Chinese or English. ** ** ### Reserved Fields To ensure your video data is processed and indexed accurately and efficiently, please pay close attention to preparing your media data according to our specific requirements before uploading. **You must use our pre\-defined field names ** **for** ** upload:** |**Field Name** |**Mandatory?** |**Field Description** | |---|---|---| |content_id|Mandatory |`*String*`|\ | | |unique identifier of the media content|\ | | |**Each data record must have a unique ID.** | |content_type |Mandatory |`*String*`|\ | | |content type, supports uploading the following enumeration values:|\ | | ||\ | | |* "**collection**": Media series object (e.g., a season of a TV drama). No video file to be uploaded for "collection" data (it will be ignored if you uploaded). Only metadata for the series is required.|\ | | |* "**video**": Video object (e.g., an individual episode of a TV drama, or a movie). Video files and their corresponding metadata to be uploaded for "video" type data. | |video_url |**Mandatory** for type **"video"** ; |`*Array* `|\ | | |A list of video download links accessible to Volcano Engine, used for video understanding and processing. If multiple `video_url` links are uploaded under the same `content_id`, system will splice them into a single video according to the list order.|\ | |Must be **empty** for type **"collection"** |:::tip \*Note|\ | | |This field supports a list of publicly accessible video download links. Crawling videos from general web links is not supported.|\ | | |Before uploading, **ensure the total size of all video files in the list does not exceed 10GB and the total duration does not exceed 3 hours.** |\ | | |Ensure these video URLs are accessible to Volcengine’s official interface domain (volcengine.com).|\ | | |**The following video file formats are supported for upload:** |\ | | ||\ | | |* Common playback formats: mp4, mkv, avi, mov, wmv, asf, rmvb|\ | | |* Streaming formats: flv, f4v, ts, mpegts, m4s, webm, m3u8|\ | | ||\ | | |:::| |parent_content_id |Must be **empty** for type **"collection"** ;|`*String*`|\ | | |If your media information has a hierarchical structure (e.g., TV dramas), use this field to identify the parent content of the video. For "video" type data, you can pass the `content_id` of the series which the video belongs to. *The * `*parent_content_id*` * of the child content is the * `*content_id*` * of the parent.* |\ | |Optional for type "video"; |**Note**: Hierarchical relationships beyond 2 levels are currently not supported. | |sequence_index |Optional for type "video" |`*Integer*`|\ | | |Indicates the sequence of multiple child content within the parent content (e.g., episode numbers of a TV drama), enabling cross\-video understanding within the same media series.|\ | | |**Note**: Cross\-video understanding will not be performed if this is not provided. | Please carefully review your data before uploading to ensure that your dataset includes all mandatory fields mentioned above and that these fields use the prescribed names. This ensures smooth data ingestion and the normal operation of subsequent features. In addition to the fixed fields, you may use **custom fields** to upload other metadata (e.g., content title, video duration, video language, etc.). You can assign any name to these fields—such as the existing names in your database—to simplify the integration process. ## Start Creation This initial step guides you to the dataset creation page to begin the workflow. 1. In the AI Search console, navigate to the **Datasets** page. 2. Click **`Create Dataset`**. In the pop\-up window, select **Item Dataset** and click **`Create Now`** to officially enter the creation process. ## Upload Data ![图片](https://p9-arcosite.byteimg.com/obj/tos-cn-i-goo7wpa0wc/51ad0bb1ab0044a2bf8a804c4e862f73 =2478x) In this step, we will guide you through naming your new video dataset and uploading a sample file to parse data fields and generate configurations. 1. Enter the **Dataset Name** on the creation page. 2. Select **Video Data** as the Dataset Type. 3. **Upload Data:** Upload a sample of your data here. This helps the system automatically identify fields and generate initial configurations. This is for configuration preview only; you can decide later whether to formally add this data to the dataset. The system supports two upload methods; toggle below the upload control to choose your preferred method. :::tip \*Note * **Mandatory Fields:** Each data record must include the following information using strictly prescribed field names: * Unique Identifier (`content_id`): A `String` type unique ID field serving as the retrieval foundation. * Content Type (`content_type`): Supports enumerated values: `"collection"` (media collection) and `"video"` (video content). * **Conditional Mandatory Fields:** These are not required in all cases, but if your data contains this information, you **must** use our prescribed field names: Video URL (`video_url`), Parent Unique ID (`parent_content_id`), and Child Content Sequence (`sequence_index`). * **Schema Consistency:** Conflicting formats for fields with the same name across multiple records will cause an error. * **Multi\-level Structure:** Supports nested objects up to 3 levels for effective indexing. * **Image Resource Handling:** Supports passing images via URL references. You can provide an image link via a field named `cover_image_url`. We recommend keeping image sizes under **5MB**. ::: ```mixin-react return ( ); ``` 4. You can choose whether to enable the ***Automatically delete video files after 7 days*** feature. If enabled, the system will automatically delete the original video files 7 days after processing and intelligent understanding is complete. This is designed to enhance data security and reduce potential leakage risks. 5. Click **`Next: Configure Fields`** to proceed to the data configuration section. ## Add Fields The platform will automatically identify most data fields based on your uploaded sample data. However, in certain cases, you may wish to add fields manually. The platform provides two methods for manual field addition: **Method 1**: Click the blue **`+ Add Field`** button at the bottom of the field list to add a new field. ![图片](https://p9-arcosite.byteimg.com/obj/tos-cn-i-goo7wpa0wc/058eace0b940427f9285ac03bdc47e53 =2486x) **Method 2**: For fields with a data type of **Object** or **Object Array**, you can click the **`+`** ** button** next to the field name. This allows you to add individual nested fields within that object (or within the objects in the array) and specify their names, types, and other information. :::tip \*Note The platform currently supports a maximum nested object depth of **3 levels**. ::: ![图片](https://p9-arcosite.byteimg.com/obj/tos-cn-i-goo7wpa0wc/18edb9d0ca6f416c9b03dad86bf38bb6 =2478x) ## Delete Field During the dataset structure configuration, if you find that certain fields are redundant, no longer needed, or were incorrectly identified, you can remove them from the data structure definition. Please note that deleting a field will remove all its associated configuration information. 1. In the field list, locate the field you wish to delete. Click the **`Delete`** button on the far right of the corresponding row to remove it. **Note:** Mandatory fields cannot be deleted. ![图片](https://p9-arcosite.byteimg.com/obj/tos-cn-i-goo7wpa0wc/5dfc5600c6464f709ac4deed2af1ad46 =2484x) ## Configure Data Attributes The system has automatically identified and pre\-filled the dataset description, field names, field types, and field attributes based on the uploaded sample data. Please review these auto\-generated configurations. Ensure that your field list correctly identifies and includes the two core fields: `content_id` (Unique Identifier) and `content_type` (Content Type). ![图片](https://p9-arcosite.byteimg.com/obj/tos-cn-i-goo7wpa0wc/60e4e71e5b0a4f9fb5aca32c05c27dd9 =2480x) :::warning \*Note Field attributes are used by downstream tasks and **cannot be changed** once configured. Please ensure the accuracy of the attribute settings before saving. ::: Compared to Image & Text datasets, Video datasets support a richer set of field attributes to meet the requirements of more complex business scenarios. |**Attribute** |**Mandatory?** |**attribute description** | |---|---|---| |content identifier |**Required & Fixed** |`*String*`|\ | | |corresponds to the `*content_id*` field | |video download URL |**Conditional & Fixed** |`*Array* `|\ | | |Corresponding to the `*video_url*` field | |parent content identifier |**Conditional & Fixed** |`*String*`|\ | | |Corresponding to the `*parent_content_id*` field | |content index |**Conditional & Fixed** |`*Integer*`|\ | | |Corresponding to the `*sequence_index*` field | |content title|Optional |`*String*`|\ | | |Title of the media content, **strongly recommend ** **to upload** | |media cover image url |Optional |`*String*`|\ | | |Download link for the media cover image | |media page link |Optional |`*String*`|\ | | |Media content jump link, for search list preview only | |video duration (seconds) |Optional |`*Int64*`|\ | | |Video duration (in seconds) | |video language (BCP 47)|Optional |`*Array* `|\ | | |Language tags that comply with BCP 47 standard|\ | | |For example, Chinese: zh; English: en | |release time |Optional |`*String*`|\ | | |The release or premiere time of media content, formatted in ISO 8601 (YYYY\-MM\-DD T HH:MM:SS), supporting precision to day or second | You can refer to [Field Attributes](/docs/85296/1544957#8d68ff39) for detailed explanations of these attributes. ## Complete Creation 1. Click **`Complete configuration`** to successfully create the dataset. Once the configuration is finalized and activated, the system will automatically perform semantic analysis and indexing of the data. # Synthesize New Fields with LLM (Optional) After completing the basic field configuration, we offer an advanced feature powered by AI Large Language Models: **"Synthesize New Fields with LLM."** This is designed to help you further tap into the potential of your existing data for data enrichment. This feature can generate entirely new, high\-value text fields based on your existing data. Examples include generating image/text descriptions, extracting core keywords, or creating relevant user query lists. These AI\-generated fields are integrated directly into the dataset, effectively enhancing search and autocomplete performance, thereby improving the overall discoverability and utility of your data. This is **not a required step** for creating a dataset, but rather an optional advanced optimization. If you are interested in using AI to automatically enrich your content, please refer to [Synthesize New Fields using Large Language Models](/docs/85296/1544976) for more detailed instructions and explanations. # View Dataset Once the video dataset is created, you can find it in the Dataset List. Enter the Dataset Details page to view basic information, field configurations, and operation history. 1. After successfully creating the video dataset, return to the **Datasets** page to see the newly created item dataset. ![图片](https://p9-arcosite.byteimg.com/obj/tos-cn-i-goo7wpa0wc/1f61f85581734175acb10200f8728562 =2480x) 2. In the Dataset List, locate the dataset you wish to view and click the **`View`** button in the "Action" column. ![图片](https://p9-arcosite.byteimg.com/obj/tos-cn-i-goo7wpa0wc/7579bf19e5e44081b65e32affb004018 =2486x) 3. **Dataset Details Page**: Once you enter the details page, you will see several key tabs: * **Configuration:** This tab allows you to view and adjust various configuration parameters for the current dataset. :::warning \*Note Changes made to the dataset configuration on this page require your manual confirmation and activation before syncing to the item dataset linked to your AI search app and updating the index configuration. ::: * **Data import:** On this tab, you can add more data. * For detailed data import procedures and format requirements, please refer to the [Import and Update Item Data in Real Time](/docs/85296/1544970) * **Data preview: ** This tab allows you to check the status of image/text data validation and processing. You can also view the timestamps for data imports and recent updates, as well as preview the data. * For a detailed introduction to data details, please refer to [View Data Details](/docs/85296/1749083). * **Activity history: ** You can view all key operations performed on the dataset, such as configuration changes and data import tasks. This allows you to track the modification history of the dataset. * In addition to the Dataset Details page, you can also view the history by clicking the **`Activity history`** button in the "Action" column for the corresponding dataset on the Dataset List page.

Viking AI Search