Abstract: A heat-sensing touch interface that profiles the temperature of touch data to identify user input is disclosed herein. Heat profiling touch input improves touch accuracy and allows users to interact with touch devices more naturally. Palm touches can be classified as non-user input, even when appearing as fragmented touches resembling finger touches, when thermal data associated with the palm touchpoint(s) exceeds the temperature range profile for finger touches. Moisture can be classified as non-user input when thermal data associated with the moisture touchpoint(s) is below the temperature range profile for finger touches. The temperature range of user input can be dynamically adjusted. Energy is conserved by activating or sampling a heat sensor array based on detection of touch data. Energy is also conserved by more accurately classifying touch inputs, resulting in reporting and processing fewer non-user inputs. Feedback can be provided at touchpoints classified as user input.

Abstract: Systems and methods are provided for managing the utilization and switching of modes for a multi-modal cursor. A user interface with selectable content is displayed, along with a cursor prompt for a multi-modal cursor. A traditional cursor mode icon and an artificial intelligence (AI) cursor mode icon are also displayed with user interface. The systems determine which mode of the multi-modal cursor to activate and utilize based on user input selecting either the traditional cursor mode icon or the AI cursor mode icon. User selection of the traditional cursor mode icon activates the first cursor mode and user selection of the AI cursor mode icon activates the second cursor mode.

Abstract: Techniques are disclosed for assisting users with unknown topics by automatically presenting information associated with the unknown topics to the users. In an example embodiment, an unknown topic is referred to or discussed during a conversation between multiple users. A candidate definition for the topic is determined, where the candidate definition is known by the user that used the topic. Based on a determination that the topic and the candidate definition are unknown to a second user in the conversation, the topic and the candidate definition are provided to one or more output devices for presentation to the second user.

Abstract: In various examples there is a method performed by a controller in Kubernetes cluster. The method comprises: identifying a job to be completed by the cluster, from a plurality of jobs. In response to identifying a job to be completed by the cluster, determining at least one sidecar container associated with the job. In response to identifying a job to be completed by the cluster, determining that the job has been completed by querying a Kubernetes control plane of the cluster. In response to determining that the job has been completed, triggering termination of the sidecar container.

Abstract: A collection of data that is extremely large can be difficult to search and/or analyze. Relevance may be dramatically improved by automatically classifying queries and web pages in useful categories, and using these classification scores as relevance features. A thorough approach may require building a large number of classifiers, corresponding to the various types of information, activities, and products. Creation of classifiers and schematizers is provided on large data sets. Exercising the classifiers and schematizers on hundreds of millions of items may expose value that is inherent to the data by adding usable meta-data. Some aspects include active labeling exploration, automatic regularization and cold start, scaling with the number of items and the number of classifiers, active featuring, and segmentation and schematization.

Abstract: Examples are disclosed that relate to power supply devices and methods for adjusting a length of a device cable extending from the power supply device. In one example a method comprises, at a power supply device comprising a device cable, determining a cable property of the device cable. Based at least in part on the cable property of the device cable, actuating a motor in the power supply device to change the length of the device cable extending from the power supply device.

Abstract: A data processing system that provides access to operator state information includes a plurality of operators that are configured to perform a computation with respect to data received from data sources. State information is associated with at least one of the plurality of operators. The data processing system also includes an object graph that comprises a representation of the computation, and that may dynamically change at runtime. The data processing system also includes an interface that provides access to the state information via the object graph. The data processing system also includes a query manager that is executable to process a graph query to retrieve the state information by traversing a plurality of nodes within the object graph. Temporal navigation is also supported. Thus, processing a graph query may involve navigating to a node in the object graph at a certain point in time.

Abstract: The present disclosure provides methods and apparatuses for implementing video effect addition. At a target application, an original video frame may be obtained from a video source; the original video frame may be provided to a video effect processing application, the video effect processing application being a Web application; and a processed video frame to which a video effect is applied may be obtained from the video effect processing application. At a video effect processing application, an original video frame may be obtained from a target application; a video effect may be applied to the original video frame to obtain a processed video frame; and the processed video frame may be provided to the target application.

Abstract: A technique supplements a language model with knowledge information retrieved from external sources. The technique operates by: receiving a query; receiving knowledge information based on the query; generating original model-input information that includes the query and the knowledge information; and presenting the original model-input information to the language model. The technique further includes: receiving an original response from the language model; generating a usefulness measure that identifies usefulness of the original response; and determining whether the usefulness measure satisfies a prescribed test. Upon determining that the usefulness measure does not satisfy the test, the technique includes: generating revised model-input information that includes feedback information; presenting the revised model-input information to the language model; and receiving a revised response from the language model.

Abstract: A method, computer program product, and computing system for matching database schemas to align attributes in corresponding tables. A first attribute of a source table structured according to a source schema is represented in a document format and a plurality of target tables structured according to a target schema are represented in the document format. The document format of each of the plurality of target tables are searched to retrieve a set of target tables that include second attributes corresponding to the first attribute based on semantic similarity between the first attribute and the plurality of target tables, and a set of second attributes is selected from the retrieved set of target tables that are similar to the first attribute using a large language model (LLM). A ranked list of second attributes that match the first attribute is then generated.

Abstract: Systems and methods may be used for access control. These systems and methods may include using a data processing system to access a video stream, the video stream including an image including a virtual background, segmenting the image into a foreground portion and a background portion to determine whether the foreground portion or the background portion of the image meets a threshold requirement, and outputting an alert in response to determining that the foreground portion or the background portion of the image fails to meet the threshold requirement.

Abstract: Methods, systems, and computer storage media for providing speech synthesis using a code-mixed speech engine in a speech synthesis system. A code-mixed speech engine supports generating natural and intelligible speech in a target speaker voice—for code-mixed-text of two or more languages—based on a code-mixed speech model that supports both code-mixing and cross-locale voice transfer scenarios. In operation, code-mixed training data associated with a plurality of different languages is accessed. A code-mixed speech model—associated with a training engine and an inference engine that support generating code-mixed synthesized speech—is generated. The code-mixed speech model is deployed. A request being received for synthesized speech of a speech synthesis service. An instance of code-mixed synthesized speech is generated. The instance of code-mixed synthesized speech is generated using the code-mixed speech model.

Abstract: Systems, methods, devices, and computer readable storage media described herein provide techniques for generating recommendations utilizing a heterogeneous distance function. In an aspect, a measure of relevancy between a first data item and a second data item is received. A setting of an adjustable parameter of a parameterized heterogeneous distance function is determined based on the measure of relevancy. The parameterized heterogeneous distance function comprises first and second sub-functions. The first sub-function calculates a distance between data items based on features of a first data type and the second sub-function calculates a distance between data items based on features of a second, different, data type. A recommendation system is caused to utilize the parameterized heterogeneous distance function to generate a recommendation based on received input. In a further aspect, the measure of relevancy is determined as a function of a measure of interactions and a measure of impressions.

Abstract: A system command line interface (CLI) is enhanced by software development tool autoloading functionality. A textual command is routed to a development tool autoloader, which maps the command to a development tool and ascertains whether a conformant version of the tool is installed. If no version is installed, or only a nonconformant version is installed, then the autoloader locates a conformant version and installs it. Then the textual command is forwarded to the installed conformant version for processing. The textual command entered in the CLI is intercepted and routed to the autoloader by a shell alias, or a search path such as an environment variable, or both. Development tool installation conformance is determined using a predefined tool configuration specification. New developers are accordingly freed from research burdens, security risks, and delays otherwise imposed by the lack of a conformant tool version on their systems.

Abstract: A computer-implemented method comprising: obtaining an asset graph representing a network, the asset graph comprising a plurality of nodes representing assets of the network and edges representing relationships between the assets; calculating, using the graph, a local connectivity metric for a node of the graph; based on the local connectivity metric, identifying a local pattern associated with a potential security risk; and identifying a candidate node or a candidate edge based on the identified local pattern associated with the potential security risk.

Abstract: Apparatus and methods for prime field modular reduction are described. As an example, a custom modular reduction digital circuit for reducing an n-bit integer based on a modulus, where the modulus comprises a k-bit integer for use with a cryptographic algorithm, is described. The custom modular reduction digital circuit includes a first circuit to generate at least two partial results by processing: (1) k lower order significant bits of the n-bit integer and (2) at least a subset of bits for congruent representations corresponding to any n-k higher order bits of the n-bit integer that are higher in significance than the most significant bit of the k-bit integer. The custom modular reduction digital circuit further includes a second circuit to process the at least two partial results, output by the first circuit, to generate a reduced version of the n-bit integer for use with the cryptographic algorithm.

Abstract: Systems, methods, apparatuses, and computer program products are disclosed for determining network topology information using domain name system (DNS) queries. Network connection and dependency information of elements in a compute cluster are determined from DNS requests. A network topology is generated based on the determined network connection and dependency information. A network policy is generated for the computing cluster based on the network topology.

Abstract: Polynucleotides used for encoding information are synthesized with universal base analogs that participate in pi-stacking interactions but do not form Watson-Crick hydrogen bonds with other bases. The universal base analogs may have pyrrole-based bases such as 5-nitroindole (5NI). Inclusion of the universal base analogs in the polynucleotides prevents polymerase-based amplification such as PCR. However, the non-amplifiable polynucleotides are able to hybridize to complementary strands and the sequences may be read by nanopore sequencing. The polynucleotides may be used as molecular taggants to label items for the prevention of forgery. The ability of polynucleotides collected from an item to hybridize with known sequences can be used to establish authenticity of an item. Alternatively, the polynucleotides may be used to encode digital data in a read-only molecule that cannot be readily copied.

Abstract: The presently disclosed magnetic locking mechanism(s) for a computing device enclosure is directed at providing a fast, but tamper resistant solution for assembly and disassembly of a computing device having top and base enclosures that come together to form an overall enclosure for the computing device. Top and base enclosures that incorporate one or more of the presently disclosed magnetic locking mechanisms are capable of being quickly and easily attached and detached without damaging the computing device, so long as a correct magnetic key(s) are used. This aids both repairability and upgradability of the computing device during its life cycle, as well as recyclability at the end of its life cycle. Without the correct magnetic key(s), it is difficult to separate the enclosures without damaging one or both of the enclosures.

Abstract: A microscope system includes a laser light source to emit illumination light toward an observation sample. A plurality of optical fibers are disposed along an optical path between the laser light source and the observation sample, such that each optical fiber of the plurality of optical fibers propagates illumination light toward the observation sample. Each optical fiber of the plurality of optical fibers is longer than a coherence length of the illumination light. A first optical fiber of the plurality of optical fibers has a first length that differs from a second length of a second optical fiber by at least the coherence length.

Abstract: A technique suggests resources within a particular environment. The technique operates by receiving a query via a search interface presentation, and using a prefix tree data structure to retrieve candidate resource descriptors that match the query. The candidate resource descriptors are data items that describe different types of candidate resources. The technique further includes: generating feature information that describes the candidate resource descriptors; using a machine-trained ranking model to generate scores associated with the candidate resource descriptors, based on the feature information; sorting the candidate resource descriptors based on the scores, to provide sorted resource descriptors; and generating output information that represents a selected set of resource descriptors drawn from the sorted resource descriptors, for presentation via the search interface presentation.

Abstract: Technology is disclosed for programmatically determining an event of interest that is specific to a user, and generating, based on user-meeting data, an enriched playback timeline that includes the event of interest on a graphical user interface (GUI). To determine the event of interest, this disclosure provides technologies to determine one or more meeting data features based on meeting data associated with a meeting. Based on the one or more meeting data features, this disclosure includes determining an event of interest. The event of interest may include, for example, an indication of whether a person was mentioned, an indication of whether a question was asked, an indication of whether a keyword was mentioned, an indication of whether a topic was covered, and so forth. From these events of interest, a GUI that includes an enriched meeting playback timeline that includes an indication of the event of interest may be generated.

Abstract: This document relates to context-based evaluations of objects in an image repository. For example, the disclosed implementations can employ image classification and/or object detection models to filter images in an image repository to obtain a filtered subset of images. Then, a context-weighted metric can be determined for a designated object subclassification of objects in the subset. For instance, the context-weighted metric can represent the significance of the designated subclassification relative to one or more other object subclassifications from a larger object classification. The context-weighted metric can be employed for operations such as responding to queries over the image repository.

Abstract: This disclosure provides methods and apparatuses for training an acoustic model which is for implementing cross-speaker style transfer and comprises at least a style encoder. Training data may be obtained, which comprises a text, a speaker ID, a style ID and acoustic features corresponding to a reference audio. A reference embedding vector may be generated, through the style encoder, based on the acoustic features. Adversarial training may be performed to the reference embedding vector with at least the style ID and the speaker ID, to remove speaker information and retain style information. A style embedding vector may be generated, through the style encoder, based at least on the reference embedding vector being performed the adversarial training. Predicted acoustic features may be generated based at least on a state sequence corresponding to the text, a speaker embedding vector corresponding to the speaker ID, and the style embedding vector.

Abstract: A system for generating a user interface described herein can include a processor to detect a type of the system based on hardware components residing within the system or coupled to the system and determine a user interface manager to execute based on the type of the system. The processor can also execute the user interface manager to generate a user interface for the system, wherein the type of the user interface manager comprises a plurality of rules to indicate a layout of the user interface.

Abstract: A method and system for securely routing traffic in a computing environment via a firewall, the method including configuring a reverse proxy server via a first configuration file and configuring a DNS server via a second configuration file, and routing traffic via the reverse proxy server by looking up addresses in the DNS server. The first configuration file includes a plurality of target IP addresses, each of the plurality of target IP addresses referencing a DNS record in the DNS server and the second configuration file includes a plurality of DNS records, where each of the DNS records is initially set to point to a default IP address. When it is determined that there is a change to an IP address of a resource in the computing environment, the DNS record associated with the resource in the DNS server is automatically updated via an API call to the DNS server by replacing the default IP address with the updated IP address for the resource.

Abstract: Techniques and assemblies for enhanced circuit board layouts and circuit configurations are presented. In one example, an apparatus includes a circuit board having a power plane configured to carry electrical current from a power supply circuit to load circuitry. The power plane comprises a fence region that establishes an increased characteristic impedance of the power plane and has vias arrayed across the power plane which perforate the power plane and electrically couple to a ground plane of the circuit board. A filter is established at the fence region by at least capacitance elements electrically coupled between the power plane and the ground plane, the filter configured to route at least a portion of noise carried by the power plane to the ground plane at the fence region.

Abstract: Techniques for toggling a visibility of tagged spatial data and for correlating the tagged spatial data with output of an LLM are disclosed. Scene data describing a real-world scene in which a 3D object is located is accessed. A digital file that models the 3D object is accessed. The digital file includes tagged spatial data associated with the 3D object. User input is received. The user input is directed to the 3D object. The digital file and the user input are provided as input to the LLM. The LLM correlates the user input with the tagged spatial data and generates a response. While the LLM's response is being provided to the user, a display of a hologram is toggled, where this hologram overlays at least a portion of the 3D object.

Abstract: A computer system configured to thinly provision a plurality of logical volumes over a plurality of types of storage. The plurality of logical volumes includes (1) a first logical volume that is thinly provisioned over a first type of storage, and (2) a second logical volume that is thinly provisioned over a second type of storage. The status of the plurality of logical volumes is monitored to detect one or more events associated with at least one of the plurality of logical volumes. In response to detecting a particular event among the one or more events, a set of data stored in a portion of the first logical volume is selected. The selected set of data is copied to the second logical volume, and the portion of the first logical volume that stores the set of data is deallocated.

Abstract: A method enacted in a depth-imaging system comprises (a) modulating radiant output from an emitter at one or more modulation frequencies; (b) projecting the radiant output as a plurality of sparse-projection features; (c) acquiring a plurality of raw shutters on an imaging sensor array modulated at the one or more modulation frequencies; (d) triangulating a geometric estimate of depth to a subject locus reflecting a sparse-projection feature; and (e) returning a time-of-flight value of the depth based on a phase computed from the plurality of raw shutters and on the geometric estimate of the depth.

Abstract: The methods and systems improve uncorrectable error (UE) and silent data corruption (SDC) rates for memory chips and improve error correction of the memory chips. The systems include a memory bank with a plurality of memory chips in communication with a memory controller. The memory bank uses one additional memory chip that stores a bitwise parity of the data stored in the remaining memory chips of the memory bank. The parity bits are used to rebuild corrupted data when a UE occurs. The parity bits are also used to detect whether a SDC occurred in the data.

Abstract: A method of updating a pose of a plurality of joints of a kinematic tree of an articulated object is described. The method comprises receiving, for each of the joints in the kinematic tree, an initial pose. A single first embedding vector is computed by encoding the initial poses in an embedding space. For each of some but not all of the joints in the kinematic tree, a target pose is received. A single second embedding vector representing the target poses is computed in the embedding space. The first embedding vector is modified using the second embedding vector to form a third embedding vector. Decoding the third embedding vector produces the updated pose of the articulated object.

Abstract: Example solutions for ranking object detection results generate or receive a plurality of segmentation masks each corresponding to one or more images. Each segmentation mask of each plurality of segmentation masks is generated using a different object detector or setting options. A quality predictor assigns a quality score to each segmentation mask, without using ground truth for the image(s). A set (one or more, but less than all) of the highest quality scores is identified for each image. In some examples, an image processing task is performed using the segmentation masks having an assigned quality score that is within the set of highest quality scores. In some examples, only the segmentation mask having the highest quality score for an image is used in the image processing task. In some examples, a quality threshold is provided, and the segmentation masks meeting the quality threshold are used in the image processing task.

Abstract: Examples of the present disclosure describe systems and methods for. In some examples, a software agent collects data from a node, such as logs or monitoring information, and provides the data to a controller. The controller assesses the attestation state and the configuration drift of the node. In some examples, the controller applies a taint to the node, which may indicate a condition or constraint on the node. A scheduler manages the workloads on the node based on the attestation state, the configuration drift, and in some examples, the taint of the node. The scheduler decides whether to schedule a workload to the node, evict a workload from the node, or keep a workload on the node depending on the attestation state and configuration drift of the node, for example, whether the workload has a toleration for the taint of the node.

Abstract: Provided is a conformal electromagnetic interference (EMI) shielding film including a thermal-forming film layer and an electrically conductive film layer. The thermal-forming film layer is configured to conformally coat over one or more electronic components mounted on a substrate with application of heat. The electrically conductive film layer is formed on an opposite side of the thermal-forming film layer from the substrate and has a plurality of voids that are configured to deform during the application of heat and allow the electrically conductive film layer to conform together with the thermal-forming film layer.

Abstract: A method for dividing an optical fibre comprises providing a hollow core optical fibre having a length, a cladding which comprises a microstructure, a hollow core surrounded by the cladding, and a negative pressure portion in which the pressure is below atmospheric pressure that extends over at least a portion of the length; and cleaving the hollow core optical fibre at a divide location within the negative pressure portion to divide the hollow core optical fibre into two parts and provide an end facet at the divide location on one or both of the parts.

Abstract: Message transformation rules for integration platforms are represented in a computing system using transformation rule trees of objects in map object models. The transformation rule tree is an intermediate representation of the transformation rule. The objects in the tree are instantiated from classes of one or more message transformation class hierarchies. Some example classes include iterator statement, conditional statement, and expression classes. Some example expression classes include constant, variable, function, and selector classes. In a message transformation architecture, transformation rules map incoming message schemas to outgoing message schemas. Transformation engines transform incoming messages to outgoing messages by applying the map object models to select data fragments of an incoming message and construct an outgoing message, from the fragments and from computational results of routines invoked on the fragments.

Abstract: In neural-network-based approaches to sequence modeling, an output sequence may be modeled via segmentations, the probability of the output sequence being constructed as a sum of products of output-segment probabilities, taken over all valid output-sequence segmentations. A set of artificial neural networks may model the distribution of the output-sequence probability with a recurrent neural network modeling the distributions of the individual output-segment probabilities, optionally in conjunction with a second recurrent neural network modeling concatenations of output segments. In various embodiments, this approach is applied to neural phrase-based machine translation.

Abstract: Embodiments are described for automatically generating threshold values based on a target metric value that specifies a desired precision or recall performance of an ML model. For instance, a trained ML model is executed against a data set using possible threshold values. Accuracy metric(s) of the ML model is determined based on the execution. Using the accuracy metric(s), evaluation metrics are modeled. A probability that a first modeled evaluation metric value has a relationship with a target metric value is determined. A determination is made that the probability has a relationship with a confidence level. Responsive to determining that the probability has the relationship with the confidence level, the threshold value is added to a set of candidate threshold values. The threshold value from among the set of candidate threshold values is selected by selecting the candidate threshold value associated with the largest second modeled evaluation metric value.

Abstract: A computer-implemented method includes receiving a plurality of security signals relating to at least one entity. A string is generated based on the plurality of security signals, the string including a name of each security signal and a timestamp of each security signal. The string is input to a trained embedding model and in response an embedding vector corresponding to the string is received. The embedding vector is input to a trained machine learning model and in response an output is received, the output being indicative of whether the plurality of security signals is representative of malicious activity.

Abstract: A neural classifier model is used to detect cybersecurity vulnerabilities in the source code predicted by a deep learning code generation model having been trained on source code possibly containing security bugs. Upon the classifier model classifying a given source code snippet as likely containing a cybersecurity vulnerability, a proposed repair for the cybersecurity vulnerability is predicted from a neural decoder transformer model having been trained on non-vulnerable source code. The neural decoder transformer model is used to predict source code that repairs the cybersecurity vulnerability given the source code classified with a cybersecurity vulnerability.

Abstract: Examples of the present disclosure describe systems and methods for automatically detecting and repairing reliability issues in operating systems and applications using a generative artificial intelligence (“AI”) system. In examples, a generative AI system receives a request to evaluate a detected issue in a software service or application. In response, the system analyzes error information associated with the detected issue to build an error context for the software code that caused the detected issue. The error context is used to identify the location of the software code file that comprises the software code. The error context and the software code file are used to identify a prompt. The prompt, the error context, and/or the identified software code are provided as input to a language model. The language model provides an output that is responsive to the user request and may perform actions to further evaluate or repair the detected issue.

Abstract: Systems, methods, and techniques described herein relate to forecasting power consumption for maintenance of data centers. In an aspect, a maintenance window request for a requested maintenance task with respect to a data center is received. A machine learning (ML) model utilized to generate a window recommendation result, the model trained to determine a period of time where the task's impact on performance of the data center satisfies an impact criterion. The window recommendation result comprising a window indicating a period of time wherein impact of the requested maintenance task satisfies the impact criterion. Performance of the maintenance task is to be performed during the window. In an alternative aspect, the ML model is trained to forecast power consumption of a data center in a period of time. In this alternative, a maintenance window determiner determines the window recommendation result based on a power consumption forecast of the ML model.

Abstract: Images captured by a camera moving in an environment are received, and for each of a plurality of points in one of the images, outputs are computed using a neural network. The outputs comprise: a trajectory depicting the point in each of the plurality of images, as well as, for each trajectory, a prediction of visibility of the trajectory in each of the images and a prediction of whether the trajectory depicts a static or moving surface in the environment. The neural network receives the images and points as input and computes the outputs, wherein the outputs comprise for each of the trajectories, confidence data. The outputs are sent to a downstream process selected from any of: visual odometry, structure from motion, human body tracking, video editing, vehicle tracking.

Abstract: This document relates to processing of source code using generative language models. One example method includes accessing source code, processing the source code to identify entities in the source code, and generating a graph having nodes representing the entities in the source code and edges representing relationships among the entities. The example method also includes prompting a generative language model to generate augmentation data for the entities based at least on the relationships, receiving, from the generative language model, generated augmentation data, and generating an augmented graph by associating the generated augmentation data with respective nodes of the graph. The augmented graph provides a basis for subsequent operations on the source code by the generative language model.

Abstract: The present disclosure provides systems and methods for user interaction and task completion using multiple devices. A set of devices may be used to perform a task, such that different devices may perform different steps of the task. A device management service may update state information at each device of the set, thereby enabling a user to interact with any of the computing devices to perform the task. A device management service may also automatically determine which device should be used by the user, based on task or step requirements, device characteristics, and device capabilities, among other examples. Thus, rather than being required to continue a task on the same device (even when the device is not well-suited for the current step or task), the user is provided with the option to use and, in some instances, is automatically transitioned to use, different devices within the set.

Abstract: The description relates to resource aware object detection for encoded video streams that can identify frames of the video stream that include an object of interest, such as a human, without decoding the frames.

Abstract: The description relates to thermal management and ensuring adequate cooling of individual computing devices where multiple computing devices operate in proximity to one another. One example can obtain sensed conditions within a rack containing multiple computing devices. The rack receives cooling air from a centralized cooling system. The example can control delivery of make-up cooling air to individual computing devices within the rack to augment the cooling air from the centralized cooling system to remove thermal loads from the individual computing devices.

Abstract: Various embodiments discussed herein are directed to improving existing technologies by providing a corpus data supplement as input into a model, such as a Large Language Model (LLM). Consequently, the model can generate accurate scores or data for predictions because the model is better able to distinguish between a general understanding of natural language concepts and domain-specific concepts.

Abstract: This document relates to processing of source code using generative language models. One example method includes accessing source code, processing the source code to identify entities in the source code, and generating a graph having nodes representing the entities in the source code and edges representing relationships among the entities. The example method also includes prompting a generative language model to generate augmentation data for the entities based at least on the relationships, receiving, from the generative language model, generated augmentation data, and generating an augmented graph by associating the generated augmentation data with respective nodes of the graph. The augmented graph provides a basis for subsequent operations on the source code by the generative language model.