Knowledge base

For example: "Help me choose the Alibaba Cloud Model Studio phone with the best camera for under 3,000 CNY."

• Basic document Q&A: Ideal for semantic search on plain-text documents.
• Rich-text Reply: Ideal for use cases that require responses containing both text and images.

The knowledge base type cannot be changed after the knowledge base is created.

• Document search
  • Use cases:
    • Ideal for retrieving information from unstructured data, such as internal corporate documents and product manuals. Unstructured data is information that is not organized in a predefined data model or schema and can include text, tables, and images.
    • If your files contain images that you want the application to include in its responses, select document search.
  • Data source: You can upload local files or import them from Object Storage Service (OSS).
    Document search creation

    1. Select data: Specify a data source (files or content) for the knowledge base. The system imports the source content into the knowledge base for retrieval. You can use local upload or cloud import (by selecting an existing category or file).
       • Local upload: Upload files directly from your computer. Expand the collapsible panel below to learn how to select a parsing method.
         Parsing methods

         Configure the parsing strategy based on your needs. If you are unsure which option to choose, we recommend using the default settings.

         • Digital Parsing: Does not parse illustrations or charts in files. This is the fastest parsing method. A 10- to 20-page plain-text document is typically parsed within a few seconds to 1 minute.
         • Intelligent Document Parsing: Recognizes and extracts text from illustrations in your files to generate text summaries. These summaries, along with other non-image content, are then chunked and vectorized for retrieval. This method is relatively fast. A 10- to 20-page document with illustrations typically takes 1 to 5 minutes to parse.
         • LLM Parsing: Enables applications that use the Qwen-VL model to answer questions about the content of illustrations and charts in your files. To enable the recognition and understanding of this content, select LLM Parsing. Because this method calls an LLM for deep understanding, a 10- to 20-page document with charts typically takes 2 to 10 minutes to parse.
         • Qwen-VL parsing: Designed specifically for image files. You can specify a Qwen-VL model and provide a prompt to guide the recognition and extraction of the image layout and elements. A single image is typically parsed within a few seconds to 1 minute.
       • Cloud import: Import existing files from Object Storage Service (OSS).
    2. Index configuration: Defines how imported data is processed and stored, which directly affects retrieval performance.

       Among the following settings, only vector storage with AnalyticDB for PostgreSQL (ADB-PG) may incur fees. All other settings are free.

       • Metadata extraction
       • Excel header assembly
       • Chunking method
       • Multi-turn conversation rewriting
       • Embedding model
       • Reranking model
       • Similarity threshold
       • Maximum recall count
       • Vector storage

       Metadata consists of additional attributes for unstructured data, which are integrated into chunks as key-value pairs.

       • Purpose: Metadata provides important context for chunks and can significantly improve the accuracy of knowledge base retrieval. For example, consider a knowledge base that contains thousands of product introduction files where the file name is the product name. When a user searches for "functional overview of Product A," if the body of every file contains "functional overview" but none mention "Product A," the knowledge base might retrieve many irrelevant chunks. However, if you add the product name as metadata to all chunks, the knowledge base can accurately filter for chunks that are related to "Product A" and also contain "functional overview." This improves retrieval accuracy and reduces the LLM's input token consumption.
       • Usage: When you call an application using an API, you can specify metadata in the metadata_filter request parameter. When the application retrieves information from the knowledge base, it first filters for relevant files based on the specified metadata.
       • Note: You cannot configure metadata extraction after a knowledge base is created.

       Metadata configuration

       Enable Metadata extraction, and then click Settings to attach uniform or personalized metadata to all files in the knowledge base. During chunking, the metadata for each file is integrated into its respective chunks.

       New metadata template

       Value extraction methods

       • Constant: Attaches a fixed attribute to all files in the knowledge base.

         As shown in the preceding example, if all files in the knowledge base have the same author, you can set a constant for a field named author.

       • Variable: Attaches a variable attribute to each file in the knowledge base. The currently supported attributes are file_name and cat_name. If you select file_name, Alibaba Cloud Model Studio attaches the name of the file to its metadata, as shown in the preceding example. If you select cat_name, Alibaba Cloud Model Studio attaches the name of the category that contains the file to the file's metadata.
       • LLM: The system matches the file content against the configured Entity Description rule, extracts the relevant information, and attaches it as a metadata attribute.

         As shown in the metadata template in the preceding example, to extract all years that appear in each file as file attributes, you can configure an LLM field named date. The entity description is configured as follows:

         In the Entity Description input box, enter date information, extract only the year, then click OK.
       • RegEx: The system matches the text content of each file in the knowledge base against the specified regular expression. Content that matches the expression is extracted and added as an attribute to the file's metadata.

         As shown in the meta information template in the example above, if you need to extract all references that appear in each file (assuming that a reference is any text that starts with '《' and ends with '》'), you can configure a regular expression field named reference. The regular expression is configured as follows:

         Set the regular expression value to 《.*?》.
       • Keyword search: The system searches each file for preset keywords and adds the matched keywords as attributes to the file's metadata.

         For example, in the metadata template in the preceding example, the preset keywords are:

         Because the file contains only the keywords "financing," "industry," "green," and "capital," the system extracts only these four keywords as the value for the file's keywords attribute.

       Used for Retrieval: If enabled, the metadata field and value are used for knowledge base retrieval along with the chunk content. If disabled, only the chunk content is used for retrieval.Used for Model Reply: If enabled, the metadata field and value are provided to the LLM as input for response generation along with the chunk content. If disabled, only the chunk content is provided to the LLM for response generation.

       When enabled, the knowledge base treats the first row of all XLSX and XLS files as the header and automatically appends it to each chunk (data row). This prevents the LLM from misinterpreting the header as a regular data row.Scope: This setting applies only to XLSX and XLS table files in the knowledge base, and it treats the first row of each file as the header. It does not apply to non-table formats such as PDF, Word, or Markdown. The setting takes effect when the knowledge base is created (during document chunking) and, like the chunking strategy, cannot be changed after the knowledge base is created.Processing logic: The knowledge base chunks table files by data row: each row other than the first (header) row becomes a separate chunk. When this setting is enabled, the header row is automatically appended to every data-row chunk, so that each chunk carries both the column names (header) and the row data. This lets the LLM interpret each field correctly based on its column name, and the header row is no longer processed as a regular data row, which avoids semantic confusion during retrieval and Q&A.

       You do not need to enable this setting if the knowledge base contains files in other formats, such as PDF, or if the table files do not contain a header row.

       Select smart chunking (recommended).

       Purpose: A knowledge base splits files into chunks and converts them into vectors using an embedding model. The chunks and vectors are then stored as key-value pairs in a vector database. After the knowledge base is created, you can view or edit the specific content (text and images) of each chunk.

       Chunk title: In addition to its body content, each chunk carries a title that is shown in the chunk list, used to identify the chunk, and included in retrieval. The knowledge base generates the chunk title automatically during intelligent document parsing, based on the heading hierarchy it detects. It has two forms: a short title (title, the nearest heading the chunk belongs to) and a hierarchical title (hier_title, the full path from the top-level heading down to the current heading, with levels joined by |>, for example Heading 1|>Heading 2). Therefore, the clearer a document's heading structure is (such as PDF, Word, or Markdown files with well-formed section headings), the more complete the generated chunk titles are; for table-like files that are split by data rows and usually lack a textual heading hierarchy (such as Excel), the chunk title depends on whether a heading hierarchy is detected during parsing, and may be empty when none is detected. You can also edit the chunk title manually in chunk management.

       Note: After a knowledge base is created, you cannot change the document chunking settings. An inappropriate chunking strategy may reduce retrieval and recall performance.

       When enabled, this feature uses a dedicated model to rewrite the user's query by incorporating conversation history. This process creates a complete, standalone query for knowledge base retrieval.

       An embedding model converts the original input prompt and knowledge text into numerical vectors so the system can calculate semantic similarity. The default Official Vector (text-embedding-v2) model supports multiple languages in addition to Chinese and English and normalizes the resulting vectors. This setting cannot be changed.The vector dimensions for are fixed and cannot be modified:

       • Official Vector (text-embedding-v2): 1,536 dimensions
       • qwen3-vl-embedding: Automatically enabled when the visual understanding use case is selected. It supports generating vectors for images and rich-text documents after visual understanding.

       A reranking model is external to the knowledge base. It reranks the candidate chunks from the initial vector retrieval and returns the top K chunks with the highest similarity scores. The recommended official reranker, qwen3-rerank (hybrid), considers both semantic relevance and text-matching features (such as BM25 scores) to better handle queries that require precise keyword hits. If you only need semantic ranking, select qwen3-rerank.Reranking model modeWhen you create a knowledge base, you can select one of the following three modes for the reranking model:

       • Q&A mode (default): Scores based on the "Q&A relevance" between the query and the candidate chunks. This mode is suitable for scenarios where users ask a complete question and expect to find the answer within a chunk.
       • Similarity mode: Scores based on the "semantic similarity" between the query and the candidate chunks. This mode is suitable for scenarios where the query and the chunks have a similar style of expression.
       • Custom advanced mode: Allows you to enter a natural language instruction of up to 200 characters to influence the reranking process. This mode is suitable for scenarios with special ranking requirements.

       You can select a reranking model mode only when you create a knowledge base. You cannot change it after creation. Before you configure this setting, note the following limitations:

       • Knowledge base type limitation: This feature is applicable only to document search, Data Query, and audio/video search knowledge bases. It is not supported for Image Q&A knowledge bases.
       • Use case limitation: This feature is supported only for the Basic document Q&A and Rich-text Reply use cases. It is not supported for the visual understanding (rich-text documents) and Rapid Q&A use cases.

       This threshold sets the minimum similarity score required for a chunk to be recalled from the reranking model's results.

       This is the default similarity threshold for the knowledge base. When you associate the knowledge base with a specific Alibaba Cloud Model Studio application, you can set a separate threshold for that application, which overrides the knowledge base's default threshold.

       Lowering this threshold recalls more chunks but may include less relevant content. Conversely, raising it recalls fewer chunks and may cause relevant ones to be discarded if set too high.

       You can use hit testing to fine-tune the similarity threshold to balance recall and precision.

       The system retrieves chunks related to the input from all associated knowledge bases, reranks them using a reranking model, and selects the top K most relevant chunks to add to the LLM input as context. This K value is the maximum recall count (up to 20), and it determines the number of chunks the reranking model provides to the LLM for context.

       Increasing this value can improve the LLM's response accuracy but also increases the LLM's input token consumption.

       Select a vector database to store text vectors. The Built-in vector database is sufficient for basic knowledge base operations. For advanced capabilities such as database management, auditing, or monitoring, we recommend AnalyticDB for PostgreSQL (ADB-PG).

       When you purchase an AnalyticDB for PostgreSQL (ADB-PG) instance, you must enable Vector Engine Optimization. Otherwise, Alibaba Cloud Model Studio cannot use the instance.

    Visual understanding creation

    When you select the visual understanding (rich-text documents) use case, the knowledge base uses a multimodal embedding model to visually understand documents and preserve the original layout information, instead of using traditional chunking methods.

    File format restrictions

    In the file upload area of the Select data tab, hover over View format requirements to view the requirements.

    Index configuration differences

    The index configuration for the visual understanding use case differs from that for Basic document Q&A:

    • Embedding model: The qwen3-vl-embedding model is automatically selected and cannot be changed after creation.
    • Multi-turn conversation rewriting: Can be enabled or disabled.
    • Similarity threshold: The default is 0.20.
    • Final maximum recall count: The default is 5.
    • Chunking method: Visual understanding does not use traditional chunking methods such as smart chunking or custom chunking. Instead, it understands the entire document page based on visual indexing.

    Editing restrictions

    • The embedding model (qwen3-vl-embedding) and vector storage type (built-in) cannot be changed after creation.
    • You can change the knowledge base edition only once per day.
• Data query
  • Use cases:
    • Ideal for building Q&A systems based on structured data (data organized according to a predefined table schema), such as assistants for querying FAQs, product data, or personnel information.
    • If your data consists of complete FAQ question-and-answer pairs, select Data Query. For example, if an Excel file contains two columns, Question and Answer, a Data Query knowledge base can use the Question column for retrieval and the Answer column as context for the LLM's response.

      This column-specific processing is unavailable in the document search knowledge base type.

    • You can import multiple Excel files, but their table schemas must be identical.
  • Data source integration: You can upload local XLS or XLSX files.
    Data query creation

    1. Select data: Specify a data source (files or content) for the knowledge base. The system imports the source content into the knowledge base for retrieval. You can use local upload or cloud import.
       The data source cannot be changed after you create the knowledge base. A single knowledge base supports only one data source.
       • Local upload: Upload data tables in XLS or XLSX format from your computer. The first row must be the table header.
       • Cloud import (select data table): Select an existing data table from an Alibaba Cloud Model Studio .
    2. Index configuration: Defines how imported data is processed and stored, which directly affects retrieval performance.

       Among the following settings, only vector storage with AnalyticDB for PostgreSQL (ADB-PG) may incur fees. All other settings are free.

       • Retrieval and reply settings
       • Multi-turn conversation rewriting
       • Embedding model
       • Reranking model
       • Similarity threshold
       • Maximum recall count
       • Vector storage

       • Used for Retrieval: If enabled, the knowledge base performs retrieval on this column.
       • Used for Model Reply: If enabled, retrieval results from this column are provided to the LLM as context for response generation. For example, if you enable Used for Retrieval for the "Name," "Gender," "Position," and "Age" columns, but enable Used for Model Reply only for the "Name" and "Position" columns, the knowledge base retrieves from all four columns. However, only the content from the "Name" and "Position" columns of the retrieved data is provided to the LLM as context for its response. Because the "Age" column is not enabled for model responses, the LLM associated with this knowledge base cannot answer the question "What is Zhang San's age?".

       When enabled, this feature uses a dedicated model to rewrite the user's query by incorporating conversation history. This process creates a complete, standalone query for knowledge base retrieval.

       An embedding model converts the original input prompt and knowledge text into numerical vectors so the system can calculate semantic similarity. The default Official Vector (text-embedding-v2) model supports multiple languages in addition to Chinese and English and normalizes the resulting vectors. This setting cannot be changed.The vector dimensions for are fixed and cannot be modified:

       • Official Vector (text-embedding-v2): 1,536 dimensions
       • qwen3-vl-embedding: Automatically enabled when the visual understanding use case is selected. It supports generating vectors for images and rich-text documents after visual understanding.

       A reranking model is external to the knowledge base. It reranks the candidate chunks from the initial vector retrieval and returns the top K chunks with the highest similarity scores. The recommended official reranker, qwen3-rerank (hybrid), considers both semantic relevance and text-matching features (such as BM25 scores) to better handle queries that require precise keyword hits. If you only need semantic ranking, select qwen3-rerank.Reranking model modeWhen you create a knowledge base, you can select one of the following three modes for the reranking model:

       • Q&A mode (default): Scores based on the "Q&A relevance" between the query and the candidate chunks. This mode is suitable for scenarios where users ask a complete question and expect to find the answer within a chunk.
       • Similarity mode: Scores based on the "semantic similarity" between the query and the candidate chunks. This mode is suitable for scenarios where the query and the chunks have a similar style of expression.
       • Custom advanced mode: Allows you to enter a natural language instruction of up to 200 characters to influence the reranking process. This mode is suitable for scenarios with special ranking requirements.

       You can select a reranking model mode only when you create a knowledge base. You cannot change it after creation. Before you configure this setting, note the following limitations:

       • Knowledge base type limitation: This feature is applicable only to document search, Data Query, and audio/video search knowledge bases. It is not supported for Image Q&A knowledge bases.
       • Use case limitation: This feature is supported only for the Basic document Q&A and Rich-text Reply use cases. It is not supported for the visual understanding (rich-text documents) and Rapid Q&A use cases.

       This threshold sets the minimum similarity score required for a chunk to be recalled from the reranking model's results.

       This is the default similarity threshold for the knowledge base. When you associate the knowledge base with a specific Alibaba Cloud Model Studio application, you can set a separate threshold for that application, which overrides the knowledge base's default threshold.

       Lowering this threshold recalls more chunks but may include less relevant content. Conversely, raising it recalls fewer chunks and may cause relevant ones to be discarded if set too high.

       You can use hit testing to fine-tune the similarity threshold to balance recall and precision.

       The system retrieves chunks related to the input from all associated knowledge bases, reranks them using a reranking model, and selects the top K most relevant chunks to add to the LLM input as context. This K value is the maximum recall count (up to 20), and it determines the number of chunks the reranking model provides to the LLM for context.

       Increasing this value can improve the LLM's response accuracy but also increases the LLM's input token consumption.

       Select a vector database to store text vectors. The Built-in vector database is sufficient for basic knowledge base operations. For advanced capabilities such as database management, auditing, or monitoring, we recommend AnalyticDB for PostgreSQL (ADB-PG).

       When you purchase an AnalyticDB for PostgreSQL (ADB-PG) instance, you must enable Vector Engine Optimization. Otherwise, Alibaba Cloud Model Studio cannot use the instance.

• Image Q&A
  • Use cases:
    • Ideal for building multimodal retrieval applications that support search-by-image and search-by-image-plus-text, such as product discovery assistants or visual Q&A assistants.
  • Data source integration: You can upload local XLS or XLSX files.

    XLS and XLSX files must contain publicly accessible image URLs to build image indexes. For details, see the creation instructions below.

    Image Q&A creation

    1. Select data: Specify a data source (files or content) to import into the knowledge base for retrieval. You can use local upload or cloud import (select an existing data table from a data connector).
       The data source cannot be changed after creation, and a single knowledge base supports only one data source.
       • Local upload: Upload data tables in XLS or XLSX format directly from your computer.
         • Field requirement: The data table must contain at least one field of the type image_url to generate the image index.
         • Build process: The knowledge base accesses the image URL in the image_url field, extracts visual features, and converts them into vectors for storage.
         • Retrieval process: The knowledge base compares the vector generated from the user's uploaded image with the stored image vectors and returns the most relevant records.
       • Cloud import (select a data table): Select an existing data table from your application data in Alibaba Cloud Model Studio.
    2. Index configuration: Defines how imported data is processed and stored, which directly affects retrieval performance.

       Among the following settings, only vector storage with AnalyticDB for PostgreSQL (ADB-PG) may incur fees. All other settings are free.

       • Retrieval and reply settings
       • Multi-turn conversation rewriting
       • Embedding model
       • Reranking model
       • Similarity threshold
       • Maximum recall count
       • Vector storage

       • Used for Retrieval: If enabled, the knowledge base performs retrieval on this column.
       • Used for Model Reply: If enabled, retrieval results from this column are provided to the LLM as context for response generation. For example, if you enable Used for Retrieval for the "Name," "Gender," "Position," and "Age" columns, but enable Used for Model Reply only for the "Name" and "Position" columns, the knowledge base retrieves from all four columns. However, only the content from the "Name" and "Position" columns of the retrieved data is provided to the LLM as context for its response. Because the "Age" column is not enabled for model responses, the LLM associated with this knowledge base cannot answer the question "What is Zhang San's age?".

       When enabled, this feature uses a dedicated model to rewrite the user's query by incorporating conversation history. This process creates a complete, standalone query for knowledge base retrieval.

       An embedding model converts original input prompts, knowledge text, and images into numerical vectors to enable similarity comparisons. For more information, see Text and multimodal vectorization.

       • qwen2.5-vl-embedding: Represents single-modal or mixed-modal inputs as a unified vector, suitable for cross-modal retrieval and image search. For example, if you input an image of a shirt with the text "find a similar style that looks younger," the model can fuse the image and text instructions into a single vector for understanding.
       • multimodal-embedding-v1: Generates separate vectors for each input part (image and text).
       • qwen3-vl-embedding: An upgraded version of qwen2.5-vl-embedding that further improves image-text fusion understanding and cross-modal retrieval accuracy.

       A reranking model is external to the knowledge base. It reranks the candidate chunks from the initial vector retrieval and returns the top K chunks with the highest similarity scores. The recommended official reranker, qwen3-rerank (hybrid), considers both semantic relevance and text-matching features (such as BM25 scores) to better handle queries that require precise keyword hits. If you only need semantic ranking, select qwen3-rerank.Reranking model modeWhen you create a knowledge base, you can select one of the following three modes for the reranking model:

       • Q&A mode (default): Scores based on the "Q&A relevance" between the query and the candidate chunks. This mode is suitable for scenarios where users ask a complete question and expect to find the answer within a chunk.
       • Similarity mode: Scores based on the "semantic similarity" between the query and the candidate chunks. This mode is suitable for scenarios where the query and the chunks have a similar style of expression.
       • Custom advanced mode: Allows you to enter a natural language instruction of up to 200 characters to influence the reranking process. This mode is suitable for scenarios with special ranking requirements.

       You can select a reranking model mode only when you create a knowledge base. You cannot change it after creation. Before you configure this setting, note the following limitations:

       • Knowledge base type limitation: This feature is applicable only to document search, Data Query, and audio/video search knowledge bases. It is not supported for Image Q&A knowledge bases.
       • Use case limitation: This feature is supported only for the Basic document Q&A and Rich-text Reply use cases. It is not supported for the visual understanding (rich-text documents) and Rapid Q&A use cases.

       This threshold sets the minimum similarity score required for a chunk to be recalled from the reranking model's results.

       This is the default similarity threshold for the knowledge base. When you associate the knowledge base with a specific Alibaba Cloud Model Studio application, you can set a separate threshold for that application, which overrides the knowledge base's default threshold.

       Lowering this threshold recalls more chunks but may include less relevant content. Conversely, raising it recalls fewer chunks and may cause relevant ones to be discarded if set too high.

       You can use hit testing to fine-tune the similarity threshold to balance recall and precision.

       The system retrieves chunks related to the input from all associated knowledge bases, reranks them using a reranking model, and selects the top K most relevant chunks to add to the LLM input as context. This K value is the maximum recall count (up to 20), and it determines the number of chunks the reranking model provides to the LLM for context.

       Increasing this value can improve the LLM's response accuracy but also increases the LLM's input token consumption.

       Select a vector database to store text vectors. The Built-in vector database is sufficient for basic knowledge base operations. For advanced capabilities such as database management, auditing, or monitoring, we recommend AnalyticDB for PostgreSQL (ADB-PG).

       When you purchase an AnalyticDB for PostgreSQL (ADB-PG) instance, you must enable Vector Engine Optimization. Otherwise, Alibaba Cloud Model Studio cannot use the instance.