USER MANUALS

Denodo Assistant

This page explains how to enable the Denodo Assistant features that use large language models. The page Denodo Assistant Features That Use Large Language Models lists these features.

To use these features, first you have to set up the connection to the API of the large language model (LLM) and then, activate this set of features.

LLM Configuration

In the LLM Configuration area of the Denodo Assistant section, you set up the large language model provider that the Denodo Assistant will use.

Denodo Assistant LLM provider configuration

Denodo Assistant LLM provider configuration

When setting up a provider, click Test Configuration to ensure the configuration is correct. This sends a test request to the AI service.

Denodo supports these LLM providers: Amazon Bedrock, Azure OpenAI Service, OpenAI , Gemini Enterprise Agent Platform (pka Vertex AI) and Other (OpenAI API Compatible).

Amazon Bedrock

  • Authentication. Authentication method for accessing Amazon Bedrock. Choose between AWS IAM credentials or Denodo AWS instance credentials.

    • AWS IAM credentials. Use this option if you want to specify the AWS credentials.

      • AWS access key ID. Unique identifier that specifies the user or entity making the request.

      • AWS secret access key. Secret string that is used to sign the requests you make to AWS.

      • AWS IAM role ARN (optional). IAM identity with specific permissions.

      • AWS region. The AWS region where access to the Amazon Bedrock service is available.

    • Denodo AWS instance credentials. Use this option if the AWS credentials are configured in the instance where Denodo is running.

      • AWS IAM role ARN (optional). IAM identity with specific permissions.

      • AWS region. The AWS region where access to the Amazon Bedrock service is available.

  • Context window (tokens). The maximum number of tokens allowed by the model selected.

  • Custom Temperature: Provides advanced control over the temperature parameter used in LLM requests. You can turn off temperature entirely (required for certain reasoning-focused models such as OpenAI’s o3) or specify a fixed temperature value (necessary for models like GPT-5). This option isn’t needed for the certified models in the list, as their temperature behavior is managed internally.

  • Headers. Additional headers to include in API requests. Header values can be stored encrypted.

  • Max Output Tokens: The maximum number of tokens that the model is allowed to generate in a single response.

  • Model id. Unique identifier of a model. At the moment, only Amazon Bedrock models from the Claude family of model provider Anthropic are supported. By default, some model IDs use the “us.” prefix to denote the AWS region for inference. You can replace this prefix with your region’s code if needed.

  • Use custom URL. When enabled, allows configuring a custom endpoint URL.

  • Endpoint URL. Endpoint URL for LLM requests.

For more information on the AWS authentication parameters go to the Amazon AWS security credentials reference.

Azure OpenAI

  • Azure resource name. The name of your Azure resource.

  • Azure deployment name. The deployment name you chose when you deployed the model.

  • API version. The API version to use for this operation. This follows the YYYY-MM-DD format.

  • API key. The API key.

  • Context window (tokens). The maximum number of tokens allowed by the model you deployed.

  • Custom Temperature: Provides advanced control over the temperature parameter used in LLM requests. You can turn off temperature entirely (required for certain reasoning-focused models such as OpenAI’s o3) or specify a fixed temperature value (necessary for models like GPT-5). This option isn’t needed for the certified models in the list, as their temperature behavior is managed internally.

  • Headers. Additional headers to include in API requests. Header values can be stored encrypted.

  • Max Output Tokens: The maximum number of tokens that the model is allowed to generate in a single response.

  • Specify URI: When enabled, allows configuring a custom Chat Completions URI.

  • Chat Completions URI. The custom Chat Completions endpoint.

  • Authentication. When Specify URI is enabled, allows to configure whether the custom API requires authentication with API key.

For more information on the Azure OpenAI Service API parameters go to the Azure OpenAI Service REST API reference.

When using Specify URI, we define that a specific API is eligible if it implements the official Chat Completions Azure OpenAI Service API (see https://learn.microsoft.com/en-us/azure/ai-services/openai/reference#chat-completions). Not all the parameters of the Chat Completions Azure OpenAI Service API are needed for the custom API to be compatible with the Denodo Assistant:

  • Request body. The Denodo Assistant will make requests with a request body having the following parameters: messages and temperature.

  • Response body. The Denodo Assistant needs the following parameters in the response body: id, object, created, choices and usage

OpenAI

  • API key. This is the OpenAI API key. This parameter is required.

  • Headers. Additional headers to include in API requests. Header values can be stored encrypted.

  • Organization ID. If configured, a header to specify which organization is used for an API request will be sent. This is useful for users who belong to multiple organizations. This parameter is not required.

  • Model. The model which is going to be used to generate the query. The dropdown values are the ones tested by Denodo. However, if you want to try an untested OpenAI model, you can configure it by pressing the edit icon. Note that some models shown may not work depending on your organization’s OpenAI account.

  • Context window (tokens). The maximum number of tokens allowed by the model selected.

  • Custom Temperature: Provides advanced control over the temperature parameter used in LLM requests. You can turn off temperature entirely (required for certain reasoning-focused models such as OpenAI’s o3) or specify a fixed temperature value (necessary for models like GPT-5). This option isn’t needed for the certified models in the list, as their temperature behavior is managed internally.

  • Max Output Tokens: The maximum number of tokens that the model is allowed to generate in a single response.

For more information on the OpenAI API parameters, go to the OpenAI API reference https://platform.openai.com/docs/api-reference/authentication.

Gemini Enterprise Agent Platform (pka Vertex AI) Provider

  • Project ID. The Google Cloud project ID.

  • Location. The Google Cloud region.

  • Credential file. The Google Cloud credential file. Both service-based OAuth authentication and workload identity federation (WIF) credential files are accepted.

  • Model. The Gemini model name. The drop-down menu provides a list of Denodo certified models. To use an uncertified Gemini Enterprise Agent Platform model, click the edit icon to enter a custom name manually. Ensure that any custom entry belongs to the Gemini family of models. Note that some models shown may not work depending on your Google account and organization.

  • Context window (tokens). The maximum number of tokens allowed by the model selected.

  • Custom Temperature: Provides advanced control over the temperature parameter used in LLM requests. You can turn off temperature entirely or specify a fixed temperature value. This option isn’t needed for the certified models in the list, as their temperature behavior is managed internally.

  • Max Output Tokens: The maximum number of tokens that the model is allowed to generate in a single response.

Other (OpenAI API Compatible)

  • Authentication. Configure this option depending on whether the custom API requires authentication.

  • API key. The API key. If the authentication is turned on, this parameter is required.

  • Headers. Additional headers to include in API requests. Header values can be stored encrypted.

  • Organization ID. If configured, a header to specify which organization is used for an API request will be sent. This is useful for users who belong to multiple organizations. Only available when the authentication is turned on. This parameter is not required.

  • Chat Completions URI. The Chat Completions endpoint. This parameter is required.

  • Model. The model which is going to be used to generate the query. This parameter is required.

  • Context window (tokens). The maximum number of tokens allowed by your custom model. This parameter is required.

  • Custom Temperature: Provides advanced control over the temperature parameter used in LLM requests. You can turn off temperature entirely (required for certain reasoning-focused models such as OpenAI’s o3) or specify a fixed temperature value (necessary for models like GPT-5). This option isn’t needed for the certified models in the list, as their temperature behavior is managed internally.

  • Max Output Tokens: The maximum number of tokens that the model is allowed to generate in a single response. This parameter is required.

The Other (OpenAI API Compatible) option allows the Denodo Assistant to send and process requests from APIs following the OpenAI Chat Completions API approach. We define that an API follows this approach if it implements the official Chat Completions OpenAI API (see https://platform.openai.com/docs/api-reference/chat). Not all the parameters of the Chat Completions OpenAI API are needed for the API to be compatible with the Denodo Assistant:

  • Request body. The Denodo Assistant will make requests with a request body having the following parameters: model, messages and temperature.

  • Response body. The Denodo Assistant needs the following parameters in the response body: id, object, created, choices and usage.

In addition to OpenAI models such as gpt-oss-20b, several others support the Chat Completions API, including Google’s Gemini and models that can be served using Ollama, such as Llama 3.3, Mistral, and Phi-4. The following examples demonstrate how to configure both a Gemini model and models served via Ollama.

  • Gemini Enterprise Agent Platform (pka Vertex AI). Follow these steps to use a deployed Gemini model via the Gemini Enterprise Agent Platform.

    1. Select Other (OpenAI API Compatible) as the Provider.

    2. If authentication is required, enter the API key. You can retrieve a temporary access token using the following command in Google Cloud Shell:

      gcloud auth print-access-token

    3. Specify the URI of the chat completions endpoint for the Gemini Enterprise Agent Platform API. This URI must include your specific {location} and {project} values

      https://aiplatform.googleapis.com/v1/projects/{project_id}/locations/{location}/endpoints/openapi

    4. In the Model field, enter the full name of your deployed Gemini model, such as google/gemini-2.5-pro.

    By following these steps, Denodo can securely connect to and utilize your deployed Gemini Enterprise Agent Platform model.

  • Ollama. Follow these steps to use a local LLM hosted by the Ollama server.

    Before you begin, ensure you have the following:

    1. The Ollama server is installed and running on your machine.

    2. You have pulled the model you wish to use (e.g., gpt-oss:20b or llama3.2:3b). The model must be available on your local Ollama server.

      ollama pull gpt-oss:20b

    Follow these steps to complete the configuration:

    1. Select Other (OpenAI API Compatible) as the Provider.

    2. Leave the API Key field blank, as a locally running Ollama server typically does not require authentication.

    3. Specify the URI for the Ollama chat completions endpoint. This will usually be your local machine’s address and the default Ollama port (11434), followed by the API path:

      http://localhost:11434/v1/chat/completions

    4. In the Model field, enter the name of the model you have pulled in Ollama, such as gpt-oss:20b or llama3.2:3b.

    By following these steps, Denodo can connect to your local Ollama server and utilize the selected model.

Thinking Provider Configuration

Thinking models can be configured to power the Denodo Assistant Chatbot for VQL Shell, its associated tools, or both. A thinking model is a Large Language Model (LLM) trained to “think longer” and handle complex, multi-step tasks. Before providing a final answer, it generates an internal “chain of thought.”

While this reasoning process is generally not visible to the user, it allows the model to break down problems, consider multiple approaches, and arrive at more accurate and logical conclusions.

Thinking provider configuration

Thinking provider configuration

To start using thinking models, activate the Enable thinking configuration toggle. This will reveal the provider selection and its specific parameters.

Unlike the standard provider configuration, thinking models have the following constraints:

  • Supported Providers: Only Amazon Bedrock (Anthropic Claude), Azure OpenAI Service, OpenAI and Gemini Enterprise Agent Platform (pka Vertex AI) are supported.

  • Fixed Temperature: Thinking models do not allow temperature configuration, as the creativity and internal logic are managed by the model’s own reasoning process.

When configuring a thinking model, you can select Reuse from the base LLM as provider. When this option is selected, it is not necessary to specify the provider or its credentials, only the specific model and thinking parameters are needed.

Amazon Bedrock (Anthropic Claude) Thinking Provider

The configuration for the Amazon Bedrock (Anthropic Claude) thinking provider uses the same parameters as the standard provider, with the addition of the following parameter:

  • Adaptive thinking. Determines how the model manages its reasoning process. Depending on this setting, you must configure one of the following parameters.

    • Reasoning effort. Required if Adaptive thinking is enabled. This parameter controls the amount of extended thinking the model applies to a request. The model evaluates the complexity of each prompt to determine whether and how much reasoning is necessary. At the default effort level (high), the model almost always engages in reasoning before responding. At lower effort levels, it may skip the reasoning phase entirely for simpler problems.

    • Thinking budget tokens. Required if Adaptive thinking is disabled. Specifies the maximum number of tokens that the model can allocate to its internal reasoning process.

Important

The Thinking budget tokens value must be strictly less than the Max output tokens value. Because reasoning tokens are part of the total output generated by the model, the remaining budget (Max output - Thinking budget) must be sufficient for the model to produce the final, visible response.

Azure OpenAI / OpenAI Thinking Provider

The configuration for Azure OpenAI Service and OpenAI thinking providers follows the same parameters as their respective standard providers (see Azure OpenAI or OpenAI), but includes the following reasoning-specific parameter:

  • Reasoning effort. This parameter specifies how much time and computational effort the model should dedicate to reasoning before generating a response.

Note

  • When configuring gpt-5.5 or gpt-5.4 as thinking models through OpenAI, the reasoning effort parameter is ignored.

Gemini Enterprise Agent Platform (pka Vertex AI)

The configuration for Gemini Enterprise Agent Platform (pka Vertex AI) thinking provider follows the same parameters as the standard provider.

HTTP Proxy Configuration

In the Http Proxy Configuration area of the LLM configuration section, you set up the proxy that the Denodo Assistant will use:

  • Mode. Specifies the proxy mode the Denodo Assistant will use. There are 3 modes:

    • DEFAULT. Denodo Assistant’s requests will be routed through a proxy if the vdp server proxy is activated in the section OTHERS > HTTP Proxy of Server Configuration.

    • MANUAL. The requests will use a proxy specific to the Denodo Assistant. When this option is selected, you must provide the proxy parameters specific for the Denodo Assitant’s requests.

    • OFF. The Denodo Assistant will not use a proxy.

If the proxy configured requires Basic Authorization, add this property to the JVM properties of the Virtual DataPort server and the web container:

Warning

The Gemini Enterprise Agent Platform (pka Vertex AI) provider does not support the proxy configuration.

-Djdk.http.auth.tunneling.disabledSchemes=""

For more information about configuring the JVM properties see Denodo Platform Configuration section.

General

In the Denodo Assistant > General, you configure the Denodo assistant.

Denodo Assistant configuration

Denodo Assistant configuration parameters

  • Enable Denodo Assistant features that use a large language model. This enables or disables the Denodo Assistant’s features that use a large language model.

Note

  • Enabling the Assistant allows the server to send metadata information to the configured LLM when using it. Only if the data usage is enabled, data of the view is also going to be processed and sent to the LLM in order to improve the results. Make sure your organization is aware and agrees on this.

  • Enabling this feature will result in incurring expenses for requests made to an external LLM API. Make sure that the organization is fully cognizant of this financial implication prior to enabling this option.

  • Language Options. The language in which the Denodo Assistant will answer. You can choose either the server’s i18n language (see the section OTHERS > Server i18n of Server configuration) or an alternative language.

  • Use a different language for column name generation. When enabled, the Assistant will generate the column names in a different language from the one selected before. See the Suggest Field Names section for more information.

  • Use sample data. When enabled, the Assistant will send actual data of the views to the LLM (i.e. rows that the views return). This can help improve the Assistant’s responses. For this option to work, check that the Cache Module is enabled. If enabled, enter the Sample data size. This is the number of rows retrieved for sampling. The sampling process will create a sample selecting a subset of the result rows. When enabled, for the feature Suggest Field Descriptions, the suggested field descriptions will typically include real data examples of the fields of type text. Moreover, suggested view description <Suggest View Description> may sometimes also include some data samples.

  • Add extra instructions to the metadata requests. When enabled, the Denodo Assistant will take the user’s input into account if it is relevant to the functionality as additional context when suggesting View Descriptions, Field Descriptions, Field Names and View Tags (both for the view itself and its fields). This can be useful for incorporating the company’s context or business naming conventions.

Note

  • Enabling the use sample data allows the Assistant to send actual data of the view(s) used. Make sure your organization is aware and agrees on this.

  • The vectorization data sampling is executed by an internal Denodo user; therefore, views with data sources that use pass-through authentication will not include sampling data.

  • Views with obligatory fields will not include sample data. To improve the accuracy of the Denodo Assistant, manually add sample values into the field’s description.

  • Add extra instructions to the metadata requests. When enabled, the Denodo Assistant will take the user’s input into account if it is relevant to the functionality as additional context when suggesting View Descriptions, Field Descriptions, Field Names and View Tags (both for the view itself and its fields). This can be useful for incorporating the company’s context or business naming conventions.

Chatbot Configuration

The Chatbot Configuration settings allow you to customize the behavior and performance of the Denodo Assistant within the VQL Shell. You can adjust the following parameters:

  • Chatbot model. LLM used for the Chatbot. Use thinking for best performance. and base for a faster, more cost-effective (but less capable) option. To use a thinking model, ensure it is first configured in the Thinking Model Configuration.

  • Tools execution model. LLM used for the AI-driven tools in the Chatbot. Use thinking for best performance. and base for a faster, more cost-effective (but less capable) option. To use a thinking model, ensure it is first configured in the Thinking Model Configuration.

  • Execute Query tool rows limit. Maximum number of rows that the execute query tool will return to the Chatbot. This constraint ensures efficient performance and prevents the Assistant from processing excessively large datasets when analyzing view data. Setting the limit to 0 disables the execute query tool.

Vector Database Integration

In the Vector Database Integration area of the Denodo Assistant section, you configure the embeddings model and vector database. This integration generates vector representations (embeddings) of your views and stores them in the vector database, enabling the Denodo Assistant to use them later for its AI-driven features.

Denodo Assistant Vector Database Integration

Denodo Assistant Vector Database Integration

When configuring a vector database and embedding model, click Test Configuration to ensure the configuration is correct. This will embed a test document into the vector database and remove it after to verify the configuration.

The section Features that use the Vector Integration explains which features require the vector integration.

It is important to note that the vector database integration will use the collections vdp_views and vdp_test_config and none of these should be used externally since they will be modified and deleted in the process of using the vector integration features.

Index Configuration

The Index Configuration allows users to configure how Denodo handles the indexation:

  • Batch Size. Number of elements to index at once.

  • Refresh metadata interval. Interval in seconds at which the system looks for updates and refreshes the necessary metadata.

Vector Database Configuration

Virtual DataPort can store the vectorized metadata in a database from any provider from the Database Adapter list.

  • Database URI. The connection URL to the database.

  • Authentication type. Authentication type to use when connecting to de vector database. Depending on the provider selected, the supported authentication types are: login and password, login and password from password vault (single secret) and login and password from password vault (one secret per field).

  • Specify custom catalog and schema Depending on the provider selected, a custom catalog and/or schema can be specified.

Embedding Model Configuration

Denodo supports these embedding model providers: Amazon Bedrock, Azure OpenAI Service, OpenAI and Other (OpenAI API Compatible).

Azure OpenAI

  • Azure OpenAI endpoint. Endpoint for the Azure OpenAI API.

  • Azure Deployment name. Name of the deployed Azure OpenAI embedding model.

  • Authentication. Authentication method for accessing Azure. Choose between API Key or Username and password.

    • API key. The API key.

    • Username and password. Username and password for Azure identity-based authentication.

  • Headers. Additional headers to include in API requests. Header values can be stored encrypted.

  • Max tokens. Maximum tokens allowed by the deployed model.

  • Proxy Mode. Proxy to use for the embedding requests. There are 3 modes:

    • DEFAULT. Denodo Assistant’s embedding requests will be routed through a proxy if the vdp server proxy is activated in the section OTHERS > HTTP Proxy of Server Configuration.

    • MANUAL. The embedding requests will use a proxy specific to the Denodo Assistant. When this option is selected, you must provide the proxy parameters specific for the Denodo Assistant’s embedding requests.

    • OFF. The Denodo Assistant embedding requests will not use a proxy.

For more information on the Azure OpenAI Service API parameters go to the Azure OpenAI Service REST API reference.

Amazon Bedrock

  • Authentication. Authentication method for accessing Amazon Bedrock. Choose between AWS IAM credentials or Denodo AWS instance credentials.

    • AWS IAM credentials. Use this option if you want to specify the AWS credentials.

      • AWS access key ID. Unique identifier that specifies the user or entity making the request.

      • AWS secret access key. Secret string that is used to sign the requests you make to AWS.

      • AWS IAM role ARN (optional). IAM identity with specific permissions.

      • AWS region. The AWS region where access to the Amazon Bedrock service is available.

    • Denodo AWS instance credentials. Use this option if the AWS credentials are configured in the instance where Denodo is running.

      • AWS IAM role ARN (optional). IAM identity with specific permissions.

      • AWS region. The AWS region where access to the Amazon Bedrock service is available.

  • Maximum number of tokens. Maximum tokens allowed by the selected model.

  • Model name. Embedding model to use.

  • Use custom URL. When enabled, allows configuring a custom endpoint URL.

  • Endpoint URL. Endpoint URL for embedding model requests.

For more details on Amazon Titan embedding models go to the Amazon Titan embedding models reference. For more information on the AWS authentication parameters go to the Amazon AWS security credentials reference.

OpenAI

  • API key. The API key.

  • Headers. Additional headers to include in API requests. Header values can be stored encrypted.

  • Maximum number of tokens. Maximum tokens allowed by the selected model.

  • Model. Embedding model to use. The dropdown values are the ones tested by Denodo. However, if you want to try an untested OpenAI model, you can configure it by pressing the edit icon. Note that some models shown may not work depending on your organization’s OpenAI account.

  • Organization ID. Specifies OpenAI organization.

For more details, see the OpenAI embedding models documentation https://platform.openai.com/docs/guides/embeddings#embedding-models.

For OpenAI and Other (OpenAI API Compatible) embedding models, you can set up a proxy without authentication using the following JVM properties of the Virtual DataPort server and the web container:

-Dhttps.proxyHost=proxyHost.com
-Dhttps.proxyPort=1234

Gemini Enterprise Agent Platform (pka Vertex AI)

  • Project ID. The Google Cloud project ID.

  • Location. The Google Cloud region.

  • Credential file. The Google Cloud credential file. Both service-based OAuth authentication and workload identity federation (WIF) credential files are accepted.

  • Model. The Google embedding model name. The drop-down menu provides a list of Denodo certified models. To use an uncertified Gemini Enterprise Agent Platform model, click the edit icon to enter a custom name manually. Ensure that any custom entry belongs to the Google family of models. Note that some models shown may not work depending on your Google account and organization.

  • Maximum number of tokens. Maximum tokens allowed by the selected model.

Other (OpenAI API Compatible)

  • API key. The API key.

  • Headers. Additional headers to include in API requests. Header values can be stored encrypted.

  • Maximum number of tokens. Maximum tokens allowed by the selected model.

  • Model. Embedding model to use.

  • URI. Custom endpoint for OpenAI.

The Other (OpenAI API Compatible) option allows the Denodo Assistant to send and process requests from APIs following the OpenAI Create Embeddings approach. We define that an API follows this approach if it adheres to the official Create Embeddings API (see https://platform.openai.com/docs/api-reference/embeddings/create).

There are models that adhere to the OpenAI Create Embeddings approach and are not made by OpenAI. These include Ollama models, such as nomic-embed-text, mxbai-embed-large or all-minilm, which can even be run locally. To configure these models, ensure you use the v1 endpoint: /v1.

Recommendations and Limitations for the Denodo Assistant Vector Database

Recommendations and Limitations for the Vector Database

The RAM and CPU requirements for the vector database depend on your concurrent usage of Denodo Assistant features. We recommend starting with a basic setup and monitoring the database health, increasing resources as needed.

Supported database providers:

  • Microsoft SQLServer 2025 or higher

    • Limitations: It does not support embedding models with more than 1,998 dimensions (for instance, the OpenAI embedding model text-embedding-3-large is not supported).

  • Oracle 23ai or higher

  • PostgreSQL 14 or higher (we recommend 16 or higher)

    • Requirements: With the PGVector extension installed and enabled.

Vector Database Disk Usage Estimation

Disk usage varies depending on the dimensions of the selected embedding model. For instance, using the OpenAI embedding model text-embedding-3-large, which generates vectors with 3,072 dimensions, each vectorized view requires approximately 13.05 kilobytes of storage. On average, the disk space required for 100,000 views ranges between 1.20 gigabytes and 2 gigabytes, depending on the database provider.

Disk Usage Estimation (100,000 views)

Database Provider

Virtual DataPort estimated Disk Usage

Embedding Model

PostgreSQL

1.2 Gigabytes

text-embedding-3-large

Oracle

1.5 Gigabytes

text-embedding-3-large

Microsoft SQLServer

2 Gigabytes

text-embedding-3-small
Add feedback