# AI search process overview AI search includes two stages: Recall, Sorting, Rerank. **Recall:** After preprocessing the user's query, the system performs cross\-modal matching recall calculation with multimodal content in the item dataset. You can adjust the recall strategy through the following configuration options: * [Configure multimodal matching strategies](/docs/85296/1873485#e0d8eb20) * [Configure the Maximum Number of Results](/docs/85296/1873485#68ef94b8) * [Configure Prioritized Item Retrieval](/docs/85296/1873485#66d097ab) * [Configure Prioritized Item Retrieval](/docs/85296/1873485#66d097ab) **Sorting:** For the retrieved search results, you can configure whether to sort them based on item popularity： * [Sorting based on Item Popularity](/docs/85296/1873485#ef8c5c0a) **Rerank:** Sorted search results can be further re\-ranked by applying customized intervention strategies based on actual needs. You can adjust the sorted results through the following settings. **If all the following strategies are enabled, Semantic re\-ranking will be performed first, then Boost and Demotion, and finally Customized sorting rules**: * [Enable Semantic Re-ranking](/docs/85296/1873485#0832ad50) * [Configure Boost and Demotion](/docs/85296/1873485#10e7aa50) * [Configure customized Sorting rules](/docs/85296/1873485#39776d7b) In addition, you can **set strategies for search terms**, including synonym dictionaries, typo correction, and autocomplete functions for search terms. You can adjust search term strategies using the following configuration options: * [Customize Synonym groups](/docs/85296/1873485#cf896f4f) * [Search Typo Correction](/docs/85296/1873485#3f11d811) * [Search Query Completion](/docs/85296/1873485#47bd3d38) After entering the app, you can configure recall, sorting, and search query strategies on the **Experience Center \> Search Experience** page and **create multiple configuration sets simultaneously**. Comparing different strategies helps identify the best practices for real business scenarios. ![图片](https://p9-arcosite.byteimg.com/tos-cn-i-goo7wpa0wc/5b087b7da7f548a2a56ae5cf3b70653e~tplv-goo7wpa0wc-image.image =2890x) :::tip tip If multiple datasets are associated with the app simultaneously, you need to configure recall and ranking strategies separately for each dataset's index. You can select the dataset to configure from the datasets tab in the **Configuration** tab. ::: # Configure multimodal matching strategies AI Search offers two methods: Text search and Image search. ## Text search After receiving a search query, AI Search uses cross\-modal hybrid retrieval to search multiple data sources and returns results ranked by relevance. You can adjust the data modality matching weights to prioritize matching item text information or item image information. You can also adjust the semantic matching weight or keyword matching weight. You can find configuration options in the sidebar: ![图片](https://p9-arcosite.byteimg.com/tos-cn-i-goo7wpa0wc/5d591a8d039a4cadbda0851630f17f3a~tplv-goo7wpa0wc-image.image =2888x) You can choose from platform preset configurations or select custom mode. ### Preset configurations To simplify setup, three preset modes are available to match common business requirements: **Balanced mode**: uses preset balanced weights to leverage both vector search (understanding your natural language and image intent) and keyword search (precise matching of specific terms). This is the most general setting, recommended for first\-time use. **Semantic search priority:** this mode relies mainly on vector models to calculate semantic similarity. This mode works better when your search query is a natural question, descriptive long sentence, or when you want to perform "text\-to\-image" search. **Keyword search priority:** this mode relies mainly on keyword matching. This mode offers higher accuracy when you need to find content containing specific product models, proper nouns, codes, or other exact text. ### Custom configuration If the preset configuration does not meet your needs, try enabling custom configuration to adjust the matching weight for different feature types. **Weight for matching image/text content:** This determines whether a text query matches more on the information in the text or features in the image within the candidate pool. **Weight for matching text semantics/keyword content:** This determines whether matching text content focuses more on semantic features or original keywords during retrieval. These two matching methods each have their focus. Understanding their differences helps you better adjust according to actual needs. * **Semantic matching: understanding the "meaning" of content** Semantic matching is achieved through vector retrieval, encoding the "meaning" or "context" of text content into high\-dimensional vectors. It pays more attention to the **deep semantic relevance** between the query content and the stored texts in the database. Even if the query does not contain exactly the same keywords, semantic matching can find highly relevant results as long as their meanings are similar. **Applicable scenarios:** \* When you want search results matched by the **concept** or **topic** of the content. \* When user queries are very colloquial, long sentences, or include synonyms. \* When cross\-lingual search is needed or handling similar content expressed in various ways. * **Keyword matching: precise hit on "literal" content** Focus more on **exact matching** between query terms and document terms. If the queried keyword appears in the document, the document is considered relevant. **Applicable scenarios:** \* When you need to perform an exact search based on **specific terminology, product models, names, or codes**. \* In scenarios demanding high vocabulary accuracy, such as legal, medical, or technical documents. \* Users clearly know the specific keywords they are looking for. ![图片](https://p9-arcosite.byteimg.com/tos-cn-i-goo7wpa0wc/ad0011c60d4c4532b40a26d8cb2930bd~tplv-goo7wpa0wc-image.image =446x) ## Image search If the associated dataset in the application has created an image index, you can enable image search functionality. After uploading an image to initiate a search request, AI search will retrieve image information from items and return results sorted by similarity. Currently, there are two supported modes: **Visual Similarity Search**: Retrieves similar images based on the overall content of the image, rather than focusing only on specific parts of the image **Item Similarity Search**: Ignores the background and human subjects to focus specifically on **finding the similar item** shown in the image (e.g., finding the same pair of shoes from a street photo). ![图片](https://p9-arcosite.byteimg.com/tos-cn-i-goo7wpa0wc/58e107c6a784427fae2d117e3031adc7~tplv-goo7wpa0wc-image.image =2898x) # Configure the Maximum Number of Results This parameter controls the maximum number of result entries returned per search request for a **single associated dataset** when users search through this application. It directly determines the upper limit of information obtained from each data source. If your application associates multiple datasets, you can **set independent Maximum Number of Results** for each dataset. This allows differential control of result quantity based on the importance and content characteristics of each dataset. ![图片](https://p9-arcosite.byteimg.com/tos-cn-i-goo7wpa0wc/23e10d809e4747dabdf719b0e43edf22~tplv-goo7wpa0wc-image.image =2924x) **Default value**: `200` **Configuration constraint:** Allowed values range from `1` to `5000` as any integer **Example:** Suppose you set the **Maximum Number of Results** for dataset A associated with your application to `5`. Then, even if a user’s search query matches 100 results in dataset A, the system will return at most 5 of the most relevant results from dataset A. # Configure Prioritized Item Retrieval Prioritizing recall ensures that specific datasets or types of content appear first in search results, preventing high\-value or time\-sensitive items from being "buried" or filtered out during the initial retrieval phase. If your candidate items are numerous but the final recall list is relatively short (for example, a base of 100,000 products with only 200 recalled results per search), you might encounter the following issues: * Every search returns content from over a month ago. Although relevant to the user’s query, the timeliness is low. * Do you want promotional categories, brands, and other items to be exposed in the search results? * Despite boosting content with high timeliness and popularity, the first page of search results still shows old or less popular content? This happens because the default search strategy matches the query against the entire pool of candidate items, selecting the final search results based on relevance scores, then applies strategy\-based re\-ranking. Newer or more popular content may have lower relevance scores, causing them to be excluded from recall and unable to participate in subsequent boosting and re\-ranking strategies. ![图片](https://p9-arcosite.byteimg.com/tos-cn-i-goo7wpa0wc/d4ec0e5659d4475f98cb08b66d986647~tplv-goo7wpa0wc-image.image =1832x) You can configure key recall protection to create separate recall channels for specific content you want guaranteed exposure or boosted ranking, ensuring these items are recalled and included in later sorting and boosting. These key\-protected contents, according to your business strategy, can be: * New content (for example, content published within the last month) * Popular content (for example, content with likes or comments exceeding a certain threshold) * Content requiring exposure support (such as promotional product categories, promoted content tags, etc.) * ... ![图片](https://p9-arcosite.byteimg.com/tos-cn-i-goo7wpa0wc/441b29afc1ea4d378a8a9408b16ab264~tplv-goo7wpa0wc-image.image =2106x) ## Create Prioritized Item Retrieval in the console Configure Prioritized Item Retrieval in the console; after saving and activating the configuration, it applies to the corresponding search strategy API calls. In the console, under Experience Center \- Search panel, find the search configuration tab you want to configure. Under the Item Retrieval section, you can configure Prioritized Item Retrieval strategies: ![图片](https://p9-arcosite.byteimg.com/tos-cn-i-goo7wpa0wc/be4820860070498b9f9e5deb59de6a14~tplv-goo7wpa0wc-image.image =4158x) Click to add a Prioritized Item Retrieval recall channel: ![图片](https://p9-arcosite.byteimg.com/tos-cn-i-goo7wpa0wc/462dcb43e04144d4bd5f96f36fa4837e~tplv-goo7wpa0wc-image.image =2010x) Define which content this Prioritized Item Retrieval recall channel recalls by configuring item field conditions; recall conditions support complex condition combinations. Different field types support different types of filter conditions. See the table below: |**Field type** |**Supported operators** |**Input content** | |---|---|---| |**String** |is in, is not in |string list | |**integer** |greater than, less than, greater than or equal to, less than or equal to, equal to, not equal to |numeric value | |^^|within range |(minimum value boundary, maximum value boundary) | |**floating point number** |greater than, less than |numeric value | |^^|in range |(minimum value boundary, maximum value boundary) | |**string list** |contains, does not contain |string list | |**integer list** |contains, does not contain |integer list | |**boolean value** |equals, does not equal |true / false | |**date and time attribute fields** |earlier than |fixed point in time | |^^|later than |fixed point in time | |^^|dynamic time range |fixed time interval based on the system time when the API is called (the current time during the user’s search) to select a time range, such as 30 days before or after the current time. | **Configuration example:** Create a new recall channel to ensure exposure of new content. The recall conditions meet any of the following: * Update time is within the past 60 days, and the status is published * Freshness tag is 1, and the status is published ![图片](https://p9-arcosite.byteimg.com/tos-cn-i-goo7wpa0wc/2bd24088bdc644f480cdf1b95507f22c~tplv-goo7wpa0wc-image.image =5000x) You can test the effect after clicking **Confirm**. Click **Sync Changes to API** in the upper right corner; Prioritized Item Retrieval takes effect when the API is called. ![图片](https://p9-arcosite.byteimg.com/tos-cn-i-goo7wpa0wc/c08778812eb3473dbc315219c7e99f31~tplv-goo7wpa0wc-image.image =2010x) :::tip tip Prioritized Item Retrieval usage restrictions: * a maximum of three Prioritized Item Retrieval strategies can be effective under each dataset * Each Prioritized Item Retrieval strategy can configure up to 5 conditional expressions ::: # Configure Personalized Retrieval This feature adjusts the recalled results based on users’ personalized interest tags—derived from their behavioral data (typically covering dimensions such as categories and brands)—while ensuring relevance. Before enabling this feature, certain prerequisite preparations are required. ## **Prerequisites for enabling personalized retrieval** * **Upload user behavior dataset:** Personalized interest tags are inferred and computed based on behavioral data. To import behavior data, refer to [Create User Event Dataset](/docs/85296/1749126). After the dataset is uploaded, data initialization and interest tags generation may take up to **1 day**. During this period, the feature may be unavailable—please wait patiently. * **Configure interest tag field as “filterable”:** Follow the console guidance (find this feature in the console, hover over it to see details) and mark the specified fields in the item dataset as **“filterable.”** For the indexing process, refer to [Create Item Dataset - Image & Text Data](/docs/85296/1544952) Once the above configurations are completed, you can enable the personalized retrieval feature in the console. # Sorting based on Item Popularity The item popularity–based sorting feature reorders search results based on item popularity, while preserving relevance. Item popularity scores are calculated from behavioral data uploaded within your application. The more positive feedback an item receives—such as clicks, add\-to\-cart actions, purchases, reviews, favorites, and shares—the higher its popularity and ranking score. Using this feature helps users discover more popular and trending items during search. Before enabling this feature, certain prerequisite preparations are required. ## **Prerequisites for enabling popularity\-based ranking** * **Upload user behavior dataset:** Item popularity relies on aggregated behavioral data. To import behavior data, refer to [Create User Event Dataset](/docs/85296/1749126) After the dataset is uploaded, data initialization and popularity calculation will start immediately. It may take up to **1 day** to fully take effect—please wait patiently. Once data upload and computation are complete, you can turn on the item popularity sorting toggle in the console. ## Enable popularity\-based sorting On the *Experience Configuration* page, locate this feature, and switch it on to experience how it affects the search experience. After validating on the preview page, click **Update Configuration** to make the ranking strategy take effect immediately and apply it to your search service. # Enable Semantic Re\-ranking ![图片](https://p9-arcosite.byteimg.com/tos-cn-i-goo7wpa0wc/9c82c223cc9741bb92f7ec22c19bf12b~tplv-goo7wpa0wc-image.image =2890x) This area is for enabling/disabling semantic reranking and setting the number of items you want to rerank, optimizing the accuracy of your top N results. Enabling semantic reranking reduces irrelevant content interference in search results and improves result accuracy. The effectiveness of the semantic reordering feature mainly depends on the questions you ask and the related context. Enabling it will somewhat affect search results response latency. Therefore, we recommend **using Semantic Re\-ranking primarily in Conversational Search**. For regular searches, since search terms are usually short and have limited text available for semantic understanding, the benefit of **Semantic Re\-ranking** is often limited. It is recommended to **use it based on the situation**. We will **disable Semantic Re\-ranking by default** for you. If you want to enable it, you can find the setting in the search configuration on the left control panel, where you can choose whether to enable it. When Semantic Re\-ranking is enabled, the default setting is to re\-rank the top 20 results. You can also customize the number of results you want to re\-rank in the sidebar (configurable range is [2,100]). When multiple datasets are configured, enabling Semantic Re\-ranking will re\-rank results separately for each dataset. * **Note that semantic reordering only works for pure text searches.** Searches containing non\-text content such as images or videos are not supported. If you add image content to a search query / question after enabling Semantic Re\-ranking, the Semantic Re\-ranking will automatically be disabled. * If the **number of records** in your dataset or the **maximum number of results** set in your search configuration is less than the semantic reordering quantity you set, the number of reordered results will not exceed your **number of records/search result limit.** You can find more information about the corresponding API fields involved in the semantic reordering feature by querying the [search API](https://arcosite.bytedance.net/sites/1740457395001/67bd5a617a7a0b0503e30be6/editor/6805b41e6ffc5805504abd99?docLang=zh&bv=6805b41e6ffc5805504abd75) in the **SearchResult** section. # Configure Boost and Demotion You can use Boost and Demotion strategies to select certain items based on their attributes for boosting and demotion. ![图片](https://p9-arcosite.byteimg.com/tos-cn-i-goo7wpa0wc/f19cd96577c74567b499ca1ccf194ab0~tplv-goo7wpa0wc-image.image =2922x) The configuration of Boost and Demotion strategies can be abstracted as: "For items meeting {condition}, increase the ranking score by {weight} when recalled in search." Details: * {condition}: The selection criteria defining the scope of items where the strategy applies, expressed as {attribute field}:{condition operator}:{value}, for example, the "Product Name" field contains the keyword "promotion," the "Inventory" field is less than or equal to 10, or the "Price" field is greater than or equal to 1000. * {weight}: The weight value, which can be chosen between \-100% and 100% (excluding 0). You can configure boost and demotion strategies separately for each associated dataset in the application. In the search configuration interface, under the dataset tab in the configuration panel, click [Boost and Demotion] \> [Add Rule] to create a new boosting or demotion rule. :::tip tip You can configure multiple boosting and demotion rules for a single field; the weight coefficients will be added together, for example: * Boost items priced over $500 by 50% * Boost items priced over $1,000 by 60% When an item meets both conditions of price over $500 and price over $1,000, its final uplift weight is 110% (= 50% + 60%). ::: ## Boost and Demotion configuration based on attributes Currently, you can add one conditional selection for a single attribute at a time to apply promotion or demotion. The method to configure the effective scope based on item fields is shown in the table below: |**Field type** |**Supported operators** |**Accepted target values** |**Condition meaning** |**Example** | |---|---|---|---|---| |String (`String`) |equals, not equals |String (`String`) |Adjust boosting or demotion when the string field value equals or does not equal a specific string | | |List (`Array`, `Array`, `Array`, `Array`) |contains |list in the corresponding format |If the list field contains all elements in the target list, adjust to boost or demote |Items containing "suspense" in the film genre are boosted | |Integer types (`Int32`, `Int64`) |greater than, greater than or equal to, less than, less than or equal to, not equal to, equal to |numeric value |Boosting or demotion when the value in an integer field is greater than or less than a specific number |Boosting when the item's release time (UNIX timestamp) is greater than 1761661582 (2025\-10\-28 22:26:22) | |floating point number (`Float`) |greater than, less than |numeric value |Boosting or demotion when the value in a floating point number is greater than or less than a specific number |Boosting for items priced above 500 yuan | ## Boost and Demotion configuration based on geographic location When your business scenario requires dynamic adjustment of search results based on the user's geographic distance, you can achieve this by configuring promotion and demotion strategies based on geographic location. The ranking engine calculates the distance between the user’s geographic location uploaded via the interface and the item's location, then applies the promotion and demotion rules based on the calculated distance. Before setting the strategy, ensure your item data contains latitude and longitude information and is configured as longitude and latitude attributes. For details on configuring these attributes, refer to [Introduction to Image & Text Item Dataset](/docs/85296/1873471) ![图片](https://p9-arcosite.byteimg.com/tos-cn-i-goo7wpa0wc/a7b09f6e51ae4f0f921a21bdf66a09f3~tplv-goo7wpa0wc-image.image =1714x) **Configure Boost and Demotion rules in the console** ![图片](https://p9-arcosite.byteimg.com/tos-cn-i-goo7wpa0wc/4ac99efbb7964881bdd1da924e44f4b8~tplv-goo7wpa0wc-image.image =1748x) 1. Navigate to "Calculated Field" and select "Distance between User and Item". ![图片](https://p9-arcosite.byteimg.com/tos-cn-i-goo7wpa0wc/c8a91c86f6ec48229b9b965411385a71~tplv-goo7wpa0wc-image.image =400x) 2. In the scope condition settings, select "greater than" or "less than" from the dropdown, enter the distance value (e.g., 20) in the input box, and choose the unit (e.g., km). ![图片](https://p9-arcosite.byteimg.com/tos-cn-i-goo7wpa0wc/f81638eabfc94b3c8a239d7278c5e1b0~tplv-goo7wpa0wc-image.image =396x) :::tip tip When the field storing geographic locations is a JSON object list, if an item has multiple sets of "longitude" and "latitude" coordinates, for example: "location":[{"long":121.228,"lat":32.112},{"long":124.029,"lat":33.072}] When calculating the distance to the user, the platform automatically selects the set of coordinates closest to the user. ::: **Upload user location in the API request** The API request's `context` object must include the `location` field to upload the user's current geographic location during the search call. For details, see [Context](/docs/85296/1544974#7f9aa6e1). **Experience location\-based uplift search** After configuring the Boost and Demotion strategy, the Experience Center page will guide you to simulate the user's geographic location on the left to test the effects: ![图片](https://p9-arcosite.byteimg.com/tos-cn-i-goo7wpa0wc/1523c592964c4cc08baa59c36f7c17f7~tplv-goo7wpa0wc-image.image =1526x) In the "Test User" area on the right, simulate and test scenarios of user location\-based boost and demotion by specifying user's location:

|**Configuration plan** |**Description** | |---|---| |Get my location |Get your current location to simulate user testing. This requires you to grant the browser location permission.|\ | |:::tip tip|\ | |If the platform does not automatically obtain latitude and longitude after clicking "Get My Location," check:|\ | ||\ | |1. Whether the device's location feature is enabled|\ | |2. Whether the computer or browser restricts the browser’s access to location permissions|\ | ||\ | |If location access was previously denied for the browser, restart the computer to reset the permission restrictions.|\ | ||\ | |:::| |Specify user location |Manually enter latitude and longitude coordinates (e.g., `longitude: 39.9°`, `latitude: 116.4°`) | # Configure customized Sorting rules In real business scenarios, you may have sorting needs beyond query\-content relevance, such as "placing recently added content first" or "sorting products with higher recent sales first." You can use custom sorting strategies to adjust the final search result order. :::tip tip AI search supports configuring custom sorting adjustment rules in the console and also supports post\-search sorting of the search result list through the "Sort by" parameter in the Search API (see [Search - Search](https://www.volcengine.com/docs/85296/1544974) for details). Differences and applicable scenarios between the two: * **Customized sorting:** Before returning the search results, the system adjusts the ranking based on defined sorting fields. When multiple items have similar relevance scores, the system performs a local re\-ranking. The final search results comprehensively consider relevance and custom sorting rules, suitable for preset business rules like prioritizing recency or popularity. * **Post\-search sorting:** After returning the search list, globally rearrange items by the specified fields using sort by and sort order, suitable when users call the interface to reorder search results on the search list page. ::: You can configure custom sorting strategies on the Console \- Search Experience page: ![图片](https://p9-arcosite.byteimg.com/tos-cn-i-goo7wpa0wc/170ef3a79c0e41178b53875e6eb6a32b~tplv-goo7wpa0wc-image.image =2890x) You can configure item sorting fields, sorting rules, and priority among different rules as needed: ![图片](https://p9-arcosite.byteimg.com/tos-cn-i-goo7wpa0wc/26dfa4bf6c164f1eba911e11e5eb3711~tplv-goo7wpa0wc-image.image =1428x) **Select sorting fields:** You can choose a String, Float, or Int type field attribute of an item as the sorting basis, such as sales volume (Int) or listing time (a "YYYY\-MM\-DD HH:MM:SS" string). **Select ascending/descending order:** Choose ascending or descending based on the field value. Strings sort lexicographically, while numeric fields sort by mathematical value. **Adjust strategy priority:** If you use multiple strategies (e.g., descending by listing time and descending by sales volume), you need to define their priority order. You can drag the strategy items to adjust their execution order. Strategies placed on top will apply first to the search result sorting. Example: Freshness + Popularity 1. **Rule 1:** `Upload_Time` (Descending) 2. **Rule 2:** `Sales_Volume` (Descending) **The Result:** Among items with similar AI relevance, the system first sorts by `Upload_Time`. If two items were uploaded at the exact same time (e.g., August 2025), it then applies Rule 2 to place the one with higher `Sales_Volume` on top. # Customize Synonym groups AI search uses semantic vector matching. The model has learned features of consensual phrases with similar semantics, which can address general synonym matching needs to some extent (for example: "Apple Inc." and "apple"). However, for **industry\-specific terminology** or **brand\-specific vocabulary**, we recommend configuring **Custom Synonym Groups**. This ensures that your search results remain accurate and comprehensive within your particular business domain. The platform supports **synonym group** configuration. Multiple words within the same synonym group are treated as "mutual synonyms." When a user inputs any word from a synonym group, items containing other words from that group are considered relevant and will be recalled. Click [Query Configuration \- Synonyms] to add your industry or brand\-specific word clusters: ![图片](https://p9-arcosite.byteimg.com/tos-cn-i-goo7wpa0wc/9b55d8219dcb457ab283d69e0038d35f~tplv-goo7wpa0wc-image.image =2900x) Note that **a word can only exist in one unique synonym group**. Duplicate configurations confuse the system in determining the correct synonym mapping. The system will prompt a warning for duplicates upon saving: ![图片](https://p9-arcosite.byteimg.com/tos-cn-i-goo7wpa0wc/7000916f974a4af58c97170a3d47a6ae~tplv-goo7wpa0wc-image.image =4742x) After configuration, click [OK] to save to the configuration panel. After verifying on the experience page, click the [Sync Changes to API] button to activate the synonym rules immediately and apply them to the Search service. # Search Typo Correction ![图片](https://p9-arcosite.byteimg.com/tos-cn-i-goo7wpa0wc/1e8640e4a13d45409a8e9f6a5019ebca~tplv-goo7wpa0wc-image.image =2890x) Viking AI Search provides automatic keyword correction in the search box experience. When users input spelling or common errors, the system automatically identifies and suggests corrections to improve search result accuracy and user experience. Two configuration options are available under Query Configuration \> Typo Tolerance when the switch is turned on: * **Direct correction:** When the system detects spelling or common errors in search terms, it automatically corrects them to the right words and returns corrected search results directly without user confirmation. * **Suggestion only:** When the system detects errors in search terms, it searches using the user’s original terms while returning correction suggestions through the Search API. You can display a prompt at the top of the page such as *"Did you mean to search: {corrected term}?”* , allowing users to decide whether to accept the correction. Regardless of the mode selected, the Search API will always return the corrected suggestion in the response body for flexible UI integration. You can integrate this freely based on your search interface design and user experience. See details: [Search](/docs/85296/1544974). # Search Query Completion ![图片](https://p9-arcosite.byteimg.com/tos-cn-i-goo7wpa0wc/ae6896de72904bf18ccc9504c7d80f4c~tplv-goo7wpa0wc-image.image =2908x) Viking AI Search supports automatic character\-based suggestion and recommendation of related search terms in the search box experience. This boosts user input efficiency, helps users quickly find desired content, and amplifies the effectiveness and value of your search scenarios. Two configuration options are available under Query Configuration \> Autocompletion: * **Max Completion Terms:** The maximum number of suggested terms returned per request. Default is **10**. * **Minimum characters to activate completion:** This defines how many characters the user must enter before triggering a search completion request. The default is 2, meaning autocompletion and related search recommendations is not triggered when the input is less than 2 characters, but triggered when the input is 2 or more characters, as shown above. We maintain a dedicated search suggestion vocabulary for you by building a rich word index, sourced from: * **Item Dataset fields:** Values from item fields you have designated as "Autocomplete" in your dataset configuration. * **Search History:** Historically popular and high\-frequency search records within the application. You can also call this service through the search term completion API and integrate it freely according to your search interface design. See details: [Query Autocomplete](/docs/85296/1544961).