--- title: "ORCA Vision Language Models" slug: "orca-vlms" description: "Unlock efficient document processing with ORCA, Hyperscience's advanced framework for accurate information extraction from complex layouts." updated: 2026-09-14T10:17:27Z published: 2026-09-14T10:17:27Z canonical: "help.hyperscience.ai/orca-vlms" --- > ## Documentation Index > Fetch the complete documentation index at: https://help.hyperscience.ai/llms.txt > Use this file to discover all available pages before exploring further. # ORCA Vision Language Models **ORCA (Optical Reasoning and Cognition Agent)** is Hyperscience’s Vision Language Models (VLM) framework for extracting information from different documents. Unlike traditional extraction models, ORCA combines language understanding with visual reasoning, allowing it to interpret document structure and extract information from complex layouts. > [!WARNING] > **Accessing this feature** > > Your access to the feature described in this article depends on your license package and pricing plan. > > To learn which features are available to your organization and how to add more, contact your Hyperscience representative. ORCA works out-of-the-box as a pre-trained base model. However, you can also train specialized models on your own documents to improve extraction accuracy for your specific use case. In this article, you'll learn: - what ORCA is - which ORCA base models are available - how ORCA models work together - where to go next to install, configure, and train specialized models. ## Key benefits ORCA is designed for organizations that need to process different types of documents, such as structured, semi-structured, or unstructured. Using ORCA allows you to: - Reduce implementation time by using a pre-trained base model. - Process a wide variety of document types. - Improve field extraction accuracy by training specialized models on your own documents. - Extract information using both textual and visual context. - Extract tables with ORCA 2 without a layout. > [!NOTE] > VLM field and table extraction are separate capabilites > > - VLM Field Extraction model definitions and specialized-model training are field-based. Their associated Semi-structured layout must contain at least one field. > - ORCA 2 table extraction uses a separate **Table Extraction Flow** and does not require a layout. ## ORCA VLMs architecture ORCA uses a layered model architecture: - A **base model** provides general-purpose document understanding and extraction. - A **specialized** model is trained on top of a specific base model, using annotated examples specific to your use case. - A Model Definition groups the base and specialized models associated with one semi-structured layout. ![Layered ORCA architecture: a base model at the foundation, with a specialized model trained on top.](https://cdn.us.document360.io/87894cef-4958-4f3f-be6f-b75a78c82548/Images/Documentation/orca.svg?sv=2026-02-06&spr=https&st=2026-11-09T19%3A44%3A07Z&se=2026-11-09T20%3A44%3A07Z&sr=c&sp=r&sig=u756C2e%2BX0Bn1GG8CXzU%2FEG%2F5%2Bzjuksw2GkM%2Blt9Tjk%3D) > [!NOTE] > Specialized models inherit the capabilities of the base model > > A specialized model is always tied to the base model it was trained on: it cannot be transferred between base models, and a model trained on ORCA 1.0 cannot be used with ORCA 2, or the reverse. ### ORCA base models Starting in v43.2, Hyperscience supports multiple ORCA base models. The base models available in your instance depend on your license. | Base model | Available in | | --- | --- | | ORCA 1.0 | v42.0.2 and later | | ORCA 2 | v43.2 and later | #### ORCA 1.0 ORCA 1.0, introduced in v42.0.2, enables GPU-powered Vision Language Model extraction in the Hyperscience platform. It provides general-purpose field extraction capabilities and serves as the foundation for use-case-specific specialized models. ORCA 1.0: - Extracts fields from Semi-structured or unstructured documents. - Supports specialized models. - Requires a Semi-structured layout with at least one field. - Does not support out-of-the-box table extraction. #### ORCA 2 ORCA 2 is the latest generation of Hyperscience’s VLM framework. ORCA 2 introduces out-of-the-box table extraction, along with improved accuracy and automation. Unlike ORCA 1.0, ORCA 2 can extract tables without requiring a predefined Semi-structured layout or a trained Table Identification model. Extracted table data is returned in table format, with the rows and columns detected in the document. > [!NOTE] > ORCA 2 does not replace Table Identification for every use case > > Table Identification remains available when layout-specific extraction is required. Learn more in [Identification Models Overview](/v43/docs/identification-models-overview). ### Choosing a base model The capabilities available depend on the ORCA base model you're using: | Capability | ORCA 1.0 | ORCA 2 | | --- | --- | --- | | Field extraction | Yes | Yes | | Specialized models | Yes | Yes | | Out-of-the-box table extraction | No | Yes | | General prompting | Yes | Yes | > [!NOTE] > We recommend using ORCA 2 > > ORCA 2 is the recommended base model for new implementations and future specialized models. If you already use ORCA 1.0, your existing specialized models continue to work with ORCA 1.0. Existing ORCA 1.0 deployments continue to be supported. Learn more about ORCA’s general prompting capabilities in [General Prompting Block](/v43/docs/using-the-general-prompting-block). ## Migrate from ORCA 1.0 to ORCA 2 ORCA 2 is recommended for new deployments because it provides improved document understanding and out-of-the-box table extraction. Learn how to migrate in [Training a Specialized Model](/v43/docs/training-a-specialized-model). > [!NOTE] > Specialized models cannot be transferred between ORCA 1.0 and ORCA 2 > > Each ORCA base model requires its own specialized model. ## Get started with ORCA VLMs Before processing documents with ORCA, make sure your instance is configured correctly. ### Infrastructure requirements > [!NOTE] > ORCA requires a Graphics Processing Unit (GPU) and training a specialized model requires a Trainer > > ORCA requires a GPU-enabled application machine for inference. See [Infrastructure Requirements](/deployment/docs/infrastructure-requirements#gpu2) for more information. > > Training a specialized model requires a trainer machine with a GPU. See [Kubernetes Installation Overview](/deployment/docs/kubernetes-installation-overview#gpus-for-trainers). > > ORCA requires a minimum of 24 GB of VRAM for inference, and a minimum of 24 GB (48 GB recommended) for training. GPUs below these amounts are not supported. For deployment-specific instructions: - **On-premise deployments** — follow the steps outlined in “Enabling Application Machines with GPUs” in the following articles: - [Docker](/deployment/docs/enabling-application-machines-with-gpus-in-on-premise-docker-deployments) - [Podman](/deployment/docs/enabling-application-machines-with-gpus-in-on-premise-podman-deployments) - [Kubernetes](/deployment/docs/enabling-application-machines-with-gpus-in-on-premise-kubernetes-deployments). - **SaaS deployments** — contact your Hyperscience representative for more information. ### Install an ORCA base model Before using ORCA, install the required base model on your instance. For installation instructions, see [Installing ORCA VLMs](/v43/docs/installing-orca-vlms). ### Processing requirements #### Workflows To process submissions with ORCA VLMs, your implementation must use one of the supported processing flows. Depending on your use case, you can use: - **Document Processing with ORCA subflow** - **VLM QA subflow** - a **custom flow** configured to use the **ORCA blocks** required for your implementation. For information about configuring ORCA flows, see [Document Processing with ORCA Subflow Settings](/v43/docs/document-processing-flow) and [Flow Blocks](/v43/docs/flow-blocks). #### VLM Quality Assurance (QA) To generate accuracy reporting for ORCA VLM extraction, you need: - the **VLM QA Subflow** selected in the **Document** **Processing** **with** **ORCA** **subflow** - QA sample rate. > [!NOTE] > The QA subflow is not selected by default > > If QA is not configured, no QA tasks are created, and accuracy reporting is not available. Learn more in [Installing ORCA VLMs](/v43/docs/installing-orca-vlms). ![ORCA QA Sample Rate set to 100% in the flow settings.](https://cdn.us.document360.io/87894cef-4958-4f3f-be6f-b75a78c82548/Images/Documentation/ORCA 100.jpg?sv=2026-02-06&spr=https&st=2026-11-09T19%3A44%3A07Z&se=2026-11-09T20%3A44%3A07Z&sr=c&sp=r&sig=u756C2e%2BX0Bn1GG8CXzU%2FEG%2F5%2Bzjuksw2GkM%2Blt9Tjk%3D) #### Control model selection in a flow In v43.2, the **ORCA Extraction Block** includes a **Use live model** setting. It determines how Hyperscience selects the model used for extraction. - When **Use live model** is selected, Hyperscience uses the live specialized model associated with the layout. - Processing stops if no live model is available. - When **Use live model** is not selected, the base model configured in the flow controls processing. - Hyperscience uses a specialized model only if it was trained with that base model. Otherwise, it processes the document with the base model alone. Configure **Use live model** in the **Flows Studio** > **Edit Flow** > **ORCA Setting type**. ![The Use live model setting in the ORCA settings type.](https://cdn.us.document360.io/87894cef-4958-4f3f-be6f-b75a78c82548/Images/Documentation/use_live_model.png?sv=2026-02-06&spr=https&st=2026-11-09T19%3A44%3A07Z&se=2026-11-09T20%3A44%3A07Z&sr=c&sp=r&sig=u756C2e%2BX0Bn1GG8CXzU%2FEG%2F5%2Bzjuksw2GkM%2Blt9Tjk%3D) Learn more in [Document Processing with ORCA Subflow Settings](/v43/docs/document-processing-with-orca-subflow-settings). #### Flexible Extraction To review low-confidence predictions with **Flexible Extraction,** you need: - Flexible Extraction configured in the **Document Processing with ORCA Subflow**. Learn more in [Installing ORCA VLMs](/v43/docs/installing-orca-vlms#configure-orca-subflows). ![Flexible Extraction Show Machine Predictions enabled in the flow settings.](https://cdn.us.document360.io/87894cef-4958-4f3f-be6f-b75a78c82548/Images/Documentation/flex_ex.jpg?sv=2026-02-06&spr=https&st=2026-11-09T19%3A44%3A07Z&se=2026-11-09T20%3A44%3A07Z&sr=c&sp=r&sig=u756C2e%2BX0Bn1GG8CXzU%2FEG%2F5%2Bzjuksw2GkM%2Blt9Tjk%3D) #### Semi-structured layout To process documents with ORCA VLMs, you need: - a **Semi-structured** layout with at least one field. - If your implementation processes both Structured and Semi-structured documents, configure a custom flow that routes each document type to the appropriate processing path. For more information, contact your Hyperscience representative. - ORCA 2’s out-of-the-box table extraction does not require a semi-structured layout. - A release assigned to the Document Processing with ORCA subflow or any custom flows with ORCA blocks. ### General-purpose extraction After you install and configure ORCA, you can start processing documents. Because ORCA is delivered as a base model, it provides general-purpose extraction capabilities and is not adapted to your specific document types or business requirements. To adapt the base model to your use case, you must train a specialized model. ### Training a specialized model To specialize a base model for your specific business case, you need: - A model definition. - annotated training data that represents your use case. For more information, see [Model Definitions](/v43/docs/model-definitions), [TDM for ORCA VLMs](/v43/docs/tdm-for-orca-vlms), and [Training a Specialized Model](/v43/docs/training-a-specialized-model). Models that understand both the text and images in a document. They combine what’s written with where it appears on the page to help the system read and extract information more accurately. Learn more in [Visual Language Models](https://help.hyperscience.ai/latest/docs/vlm-field-extraction-models). A foundational model that provides core general-purpose capabilities and is not directly trained on customer-specific examples. Use-case specialization is achieved through additional training on top of the base model using customer-specific data. A configuration object used to manage Vision Language Models (VLMs). Model definitions define a model's scope, task, compatibility, deployment status, and associated model versions. Learn more in [Model Definitions](https://help.hyperscience.ai/latest/docs/model-definitions). A configuration within Hyperscience designed to process documents where fields and table cells are present, but their positions can vary among documents. Unlike Structured layouts, Semi-structured layouts do not rely on fixed field locations. Instead, a model is trained to find the fields and cells based on provided training examples. A machine learning model in Hyperscience that learns where cells and rows are located in Semi-structured documents. It uses examples from training to predict the position of each cell and row on a page so the system can extract the correct data. Learn more in [Identification models](https://help.hyperscience.ai/latest/docs/identification-models). A processing unit initially designed to increase the speed of graphics calculations. It is efficient in completing matrix computations, making it faster than a central processing unit (CPU) in some calculations related to machine learning. A separate machine dedicated to handling resource-heavy tasks like training Identification models. It operates independently and connects to the main application through the API. A **Workflow** or **Flow** is a customizable workflow in Hyperscience that automates the processing of documents, including steps like classification, data extraction, validation, and output. Flows streamline operations by handling tasks step-by-step with minimal manual effort. Learn more in [Flows](https://help.hyperscience.ai/latest/docs/flows). **Quality Assurance (QA)** is a process that ensures the accuracy and reliability of system outputs. In Hyperscience, QA tasks allow users to review and correct errors in classification, identification, VLM extraction and transcription. Documents may enter QA after processing through Submissions, depending on the workflow configuration. A task in Hyperscience that involves human intervention to validate or correct data extraction for Structured documents. This task is used when automatic extraction isn’t fully reliable, allowing you to transcribe or adjust specific fields to ensure accuracy. A release is a package of one or more committed layout variation versions. This collection of layouts or layout variations should reflect the various document types that the machine should expect. To use the layouts that you have created to process documents, you will need to deploy a flow that contains a release. Learn more in [Releases](https://help.hyperscience.ai/latest/docs/releases). The input used to teach machine learning models how to process documents accurately. Its structure depends on the model type: - For **Classification models**, training data consists of uploaded document pages grouped by layout. - For **Identification models**, training data includes manually annotated fields and tables to train the model to extract specific data points. ## Related - [Installing ORCA VLMs](/installing-orca-vlms.md) - [Model Definitions](/model-definitions.md) - [TDM for ORCA VLMs](/tdm-for-orca-vlms.md) - [Training a Specialized Model](/training-a-specialized-model.md)