Skip to main content

Key Terminologies used in Queryloop

A

  • Account Creation and Setup: The process of establishing a new user account and configuring initial settings, such as username, password, and email.
  • Access Control: The management of permissions and restrictions on users' access to certain parts of a system or data.
  • Alpha/Rank Ratio: A parameter in model tuning that adjusts the balance between specific components, often in fine-tuning tasks.
  • API Key: A code passed in by applications to identify the calling program, its developer, or the user making the request.
  • Authentication: The process of verifying the identity of a user, typically involving credentials like passwords or tokens.
  • Automatic Scaling: An infrastructure feature that automatically adjusts computational resources based on demand.

B

  • Batch Size: The number of samples processed in one iteration during training in machine learning.

C

  • Chain-of-Thought (CoT): A reasoning approach used in AI models to break down tasks into step-by-step processes, aiding in clarity and accuracy.
  • Chunk Size: The size of data chunks used for processing, particularly during retrieval or fine-tuning phases.
  • Chunk Window Retrieval: A retrieval technique that splits data into manageable chunks for easier processing.
  • Confidential Access: Restricted data access, reserved only for authorized users to protect sensitive information.
  • Confidentiality: Ensuring that sensitive data remains secure and inaccessible to unauthorized parties.
  • Context Coverage: The extent to which the relevant context for a given query or fact is included in retrieval or evaluation.
  • Context Justification: A reasoning provided for why a piece of retrieved information is relevant in a specific context.
  • Continuous Evaluation: An ongoing assessment of model performance to ensure quality over time.
  • Cosine Similarity: A similarity metric between two vectors, often used in text retrieval to measure document or term similarity.

D

  • Dataset: A structured set of data used for training, testing, or evaluating models.
  • Dataset Schema: The structural layout or format of data in a dataset, defining types and organization.
  • Deconstruction Retrieval: A retrieval method that breaks down complex data into smaller components for better accuracy.
  • Deployed Application: A finalized application that is running in a live environment and ready for user interaction.
  • Dot Product: A mathematical operation measuring similarity between two vectors, frequently used in retrieval models.

E

  • Email: A unique identifier required during account setup for communication.
  • Embedding Cost: The computational cost associated with creating embeddings for data.
  • Embedding Fine-tuning: Customizing an embedding model to better suit specific tasks or data.
  • Embedding Model: A model used to transform data into numerical representations, or embeddings, which are used for similarity and retrieval tasks.
  • Embedding Optimization: The process of refining embeddings to improve model performance or retrieval accuracy.
  • Epoch Count: The number of complete passes through the training dataset during model training.
  • Euclidean Distance: A distance metric calculating the straight-line distance between two points in vector space, often used in clustering.
  • Evaluation Cost: The computational or financial cost of evaluating a model’s performance.

F

  • Fact Evaluation: The process of assessing the validity and relevance of extracted facts.
  • Fact Extraction: The process of identifying and isolating key facts from data.
  • Final LLM Application: A large language model application that is fully developed and ready for deployment.
  • Fine-Tuned Model: A model that has undergone additional training on specific data to improve performance for targeted tasks.
  • Fine-tuning: The process of retraining a pre-trained model on a specific dataset to adapt it to a particular task.
  • First Name: Personal information required during account setup.

G

  • Generation Module: A component responsible for generating outputs, such as responses or predictions, in an application.
  • Golden Data: High-quality, carefully curated data used for training or evaluating model accuracy.
  • Golden Responses: Reference responses used as benchmarks for assessing the accuracy of model-generated outputs.
  • Ground Truth: The actual data used for model training or evaluation, serving as a benchmark for performance assessment.

H

  • High-Availability Hosting: A hosting configuration ensuring maximum uptime and reliability of applications.
  • Hybrid (Dense + Sparse Embeddings): A retrieval method combining dense and sparse embeddings for enhanced performance.
  • HyDE Retrieval: A hybrid dense embedding technique improving accuracy in information retrieval.
  • Hyperparameters: Parameters governing the training process in machine learning, such as learning rate or batch size.

I

  • Inference Cost: The computational or financial cost associated with running inference tasks.
  • Information Retrieval: The process of extracting relevant data from a large dataset in response to a query.
  • Input Node: The initial point in a workflow where data is introduced into a system or model.
  • Iterative Evaluation: A repetitive assessment process ensuring continuous improvement and refinement of a model.

J

  • JSON Format: A lightweight data format commonly used for structuring data in APIs and applications.

L

  • Last Name: Personal information required during account setup.
  • Learning Rate: A hyperparameter that determines the step size at each iteration during model training.
  • LLM Rerank: A technique using a large language model to reorder retrieved items based on relevance.

M

  • Maximal Marginal Relevance (MMR): A technique balancing relevance and diversity in retrieved results.
  • Metric Type: The measurement or evaluation metric used for model performance, such as accuracy or recall.

N

  • Node: A point in a network or process representing data or a task.
  • No Reranker: A retrieval method that doesn’t involve reordering results after initial retrieval.
  • NoSQL Aggregation Query: A query type used in NoSQL databases to retrieve or manipulate data.
  • Noisy Data: Data that may contain errors, inconsistencies, or irrelevant information.

O

  • Organization Name: The organization associated with a user’s account.

P

  • Paraphrasing Retrieval: A retrieval technique finding variations or paraphrased versions of queries.
  • Password: A secure string of characters used for authentication.
  • Permission Level: The access level assigned to a user, defining the actions they can perform.
  • Pinecone Storage Cost: The cost associated with storing embeddings in Pinecone, a vector database.
  • Problem Statement: A description of the problem the system or model aims to solve.
  • Prompt: A query or input provided to a model to elicit a specific response or behavior.
  • Prompt Engineering: The process of designing prompts to achieve desired outcomes from models.
  • Public Access: Data or resources that are accessible to all users without restrictions.

Q

  • QLoRA Parameters: Parameters specific to the Quantized Low-Rank Adaptation method, which enhances model efficiency.
  • Queryloop: A platform for managing and deploying AI and ML experiments.

R

  • Re-rank Method: The approach used to reorder retrieved items to improve relevance.
  • Relevance Section: The portion of context directly related to a retrieved fact or result.
  • Retrieval Accuracy: The precision or correctness of results retrieved by a model.
  • Retrieval Method: The technique or algorithm used to fetch relevant data from a collection.
  • Retrieval Module: A component responsible for handling data retrieval in an application.
  • Reranker: An algorithm or model that reorders retrieval results to enhance relevance.

S

  • Secret Access: Restricted access for sensitive data, limited to specific authorized users.
  • Self-Contained Facts: Facts that are complete and independent, not requiring external context to be understood.
  • Similarity Score: A measurement indicating the degree of similarity between two items.
  • Similarity Search: A search method used to find items similar to a query based on similarity metrics.
  • Structured Data: Data organized in a defined format, like rows and columns in a table.
  • Server Cost: The cost associated with server hosting for applications or data.

T

  • Token Generation: The process of creating tokens used for secure API access.
  • Top K: The number of top-ranking results retrieved, determined by relevance or similarity.

U

  • Unstructured Data: Data without a predefined structure, such as text, images, or audio.
  • Unstructured Data with Metadata: Unstructured data accompanied by additional context, like tags or timestamps.

V

  • Verification Code: A code sent to verify user identity during account setup or login.
  • Version Control: The management of changes to documents, code, or data, tracking updates and preserving previous versions.

Additional Terminologies

  • Document Access Control: Managing permissions for accessing specific documents, ensuring proper security.
  • Table Retrieval: A technique focused on retrieving data from tabular formats like spreadsheets or databases.
  • Similarity Metric: A measure used to quantify the similarity between two items or vectors.
  • Utterance Generation: The creation of conversational responses by a language model.
  • Query Preprocessing: The steps taken to clean, normalize, and prepare a query before feeding it into a model for processing.
  • Model Configuration: The setup of a machine learning model’s architecture, hyperparameters, and other properties before training.
  • Optimization Workflow: The series of steps taken to tune and improve the model's performance.
  • Heatmap Analysis: A visualization technique used to analyze data, often to visualize the performance of a model across different parameters.
  • Cost Breakdown: A detailed analysis of the costs associated with running machine learning tasks, such as training and inference.
  • Application Cost Analysis: The assessment of financial costs associated with the deployment and maintenance of an AI-powered application.
  • Fine-tuning Iteration: The iterative process of refining a pre-trained model to improve its performance on a specific task.
  • Row Inclusion: The process of selecting rows from a dataset based on certain criteria for inclusion in model training.
  • Schema Integration: The process of aligning and combining multiple data schemas into a unified format.
  • Data Embedding: The transformation of data into a vectorized format that captures its features for similarity tasks.
  • Vector Database: A database specifically designed for storing and retrieving vectorized representations of data.

Extended

A

  • Account Creation and Setup: The process of establishing a new user account and configuring initial settings, such as username, password, and email.
  • Access Control: The management of permissions and restrictions on users' access to certain parts of a system or data.
  • Alpha/Rank Ratio: A parameter in model tuning that adjusts the balance between specific components, often in fine-tuning tasks.
  • API Key: A code passed in by applications to identify the calling program, its developer, or the user making the request.
  • Authentication: The process of verifying the identity of a user, typically involving credentials like passwords or tokens.
  • Automatic Scaling: An infrastructure feature that automatically adjusts computational resources based on demand.
  • Activation Sparsity Control: The technique of controlling sparsity in neural network activations, often for efficiency or regularization.
  • Anchor-based Embedding: A technique in which key anchor points in the embedding space are used to guide learning or retrieval tasks.
  • Application Cost Analysis: The assessment of financial costs associated with the deployment and maintenance of an AI-powered application.

B

  • Batch Size: The number of samples processed in one iteration during training in machine learning.
  • Batchwise Contrastive Loss: A loss function applied to a batch of data, typically used in contrastive learning to compare pairs of data points.
  • Backpropagation Algorithm: A widely used optimization algorithm in training neural networks, adjusting weights based on the gradient of the loss function.
  • Batch Normalization: A technique used in training deep neural networks to normalize the input layer by adjusting and scaling activations.

C

  • Chain-of-Thought (CoT): A reasoning approach used in AI models to break down tasks into step-by-step processes, aiding in clarity and accuracy.
  • Chunk Size: The size of data chunks used for processing, particularly during retrieval or fine-tuning phases.
  • Chunk Window Retrieval: A retrieval technique that splits data into manageable chunks for easier processing.
  • Confidential Access: Restricted data access, reserved only for authorized users to protect sensitive information.
  • Confidentiality: Ensuring that sensitive data remains secure and inaccessible to unauthorized parties.
  • Context Coverage: The extent to which the relevant context for a given query or fact is included in retrieval or evaluation.
  • Context Justification: A reasoning provided for why a piece of retrieved information is relevant in a specific context.
  • Continuous Evaluation: An ongoing assessment of model performance to ensure quality over time.
  • Cosine Similarity: A similarity metric between two vectors, often used in text retrieval to measure document or term similarity.
  • Contrastive Embedding Loss: A loss function used in contrastive learning, encouraging similar data points to be closer together and dissimilar points to be farther apart in the embedding space.
  • Contextual Embedding Aggregation: The process of combining embeddings from various sources or contexts to form a unified representation.

D

  • Dataset: A structured set of data used for training, testing, or evaluating models.
  • Dataset Schema: The structural layout or format of data in a dataset, defining types and organization.
  • Deconstruction Retrieval: A retrieval method that breaks down complex data into smaller components for better accuracy.
  • Deployed Application: A finalized application that is running in a live environment and ready for user interaction.
  • Dot Product: A mathematical operation measuring similarity between two vectors, frequently used in retrieval models.
  • Dynamic Few-shot Adaption: A technique in machine learning where the model quickly adapts to new tasks with few examples, typically through transfer learning or meta-learning.

E

  • Email: A unique identifier required during account setup for communication.
  • Embedding Cost: The computational cost associated with creating embeddings for data.
  • Embedding Fine-tuning: Customizing an embedding model to better suit specific tasks or data.
  • Embedding Model: A model used to transform data into numerical representations, or embeddings, which are used for similarity and retrieval tasks.
  • Embedding Optimization: The process of refining embeddings to improve model performance or retrieval accuracy.
  • Epoch Count: The number of complete passes through the training dataset during model training.
  • Euclidean Distance: A distance metric calculating the straight-line distance between two points in vector space, often used in clustering.
  • Evaluation Cost: The computational or financial cost of evaluating a model’s performance.
  • Embedding Compression Techniques: Methods to reduce the size of embeddings, typically to enhance efficiency without sacrificing too much accuracy.

F

  • Fact Evaluation: The process of assessing the validity and relevance of extracted facts.
  • Fact Extraction: The process of identifying and isolating key facts from data.
  • Final LLM Application: A large language model application that is fully developed and ready for deployment.
  • Fine-Tuned Model: A model that has undergone additional training on specific data to improve performance for targeted tasks.
  • Fine-tuning: The process of retraining a pre-trained model on a specific dataset to adapt it to a particular task.
  • First Name: Personal information required during account setup.
  • Focal Loss in Multi-Class Embeddings: A loss function used in multi-class classification problems that focuses more on hard-to-classify examples, often used in the context of imbalanced data.

G

  • Generation Module: A component responsible for generating outputs, such as responses or predictions, in an application.
  • Golden Data: High-quality, carefully curated data used for training or evaluating model accuracy.
  • Golden Responses: Reference responses used as benchmarks for assessing the accuracy of model-generated outputs.
  • Ground Truth: The actual data used for model training or evaluation, serving as a benchmark for performance assessment.
  • Gradient-Free Optimization: A class of optimization algorithms that do not require the computation of gradients, often used when gradients are hard to compute or noisy.

H

  • High-Availability Hosting: A hosting configuration ensuring maximum uptime and reliability of applications.
  • Hybrid (Dense + Sparse Embeddings): A retrieval method combining dense and sparse embeddings for enhanced performance.
  • HyDE Retrieval: A hybrid dense embedding technique improving accuracy in information retrieval.
  • Hyperparameters: Parameters governing the training process in machine learning, such as learning rate or batch size.
  • Hierarchical Embedding Structure: An embedding structure that organizes embeddings in a hierarchy, often used for multi-level or multi-scale tasks.

I

  • Inference Cost: The computational or financial cost associated with running inference tasks.
  • Information Retrieval: The process of extracting relevant data from a large dataset in response to a query.
  • Input Node: The initial point in a workflow where data is introduced into a system or model.
  • Iterative Evaluation: A repetitive assessment process ensuring continuous improvement and refinement of a model.
  • Inverse Document Frequency (IDF) Weighting: A component of the TF-IDF algorithm, used to weigh words based on how common or rare they are across a collection of documents.

J

  • JSON Format: A lightweight data format commonly used for structuring data in APIs and applications.

L

  • Last Name: Personal information required during account setup.
  • Learning Rate: A hyperparameter that determines the step size at each iteration during model training.
  • LLM Rerank: A technique using a large language model to reorder retrieved items based on relevance.
  • Label Smoothing Regularization: A regularization technique used in classification tasks to prevent overfitting by smoothing the target labels.

M

  • Maximal Marginal Relevance (MMR): A technique balancing relevance and diversity in retrieved results.
  • Metric Type: The measurement or evaluation metric used for model performance, such as accuracy or recall.
  • Memory-Mapped File Handling: A technique for accessing data from disk that improves efficiency, often used for large datasets or in distributed settings.
  • Meta-Gradient Descent: A technique in meta-learning where the model learns to adapt its learning process over time.
  • Multi-task Learning Framework: A learning framework where a single model is trained to handle multiple tasks simultaneously, often improving performance across tasks.

N

  • Node: A point in a network or process representing data or a task.
  • No Reranker: A retrieval method that doesn’t involve reordering results after initial retrieval.
  • NoSQL Aggregation Query: A query type used in NoSQL databases to retrieve or manipulate data.
  • Noisy Data: Data that may contain errors, inconsistencies, or irrelevant information.
  • Non-Parametric Embedding Models: Embedding models that do not rely on a fixed number of parameters, allowing for greater flexibility and scalability.

O

  • Organization Name: The organization associated with a user’s account.
  • Optimization Workflow: The series of steps taken to tune and improve the model's performance.

P

  • Paraphrasing Retrieval: A retrieval technique finding variations or paraphrased versions of queries.
  • Password: A secure string of characters used for authentication.
  • Permission Level: The access level assigned to a user, defining the actions they can perform.
  • Pinecone Storage Cost: The cost associated with storing embeddings in Pinecone, a vector database.
  • Problem Statement: A description of the problem the system or model aims to solve.
  • Prompt: A query or input provided to a model to elicit a specific response or behavior.
  • Prompt Engineering: The process of designing prompts to achieve desired outcomes from models.
  • Public Access: Data or resources that are accessible to all users without restrictions.

Q

  • QLoRA Parameters: Parameters specific to the Quantized Low-Rank Adaptation method, which enhances model efficiency.
  • Queryloop: A platform for managing and deploying AI and ML experiments.
  • Query Expansion with Embeddings: A technique used in retrieval tasks where additional terms or contexts are added to a query to improve retrieval results.

R

  • Re-rank Method: The approach used to reorder retrieved items to improve relevance.
  • Relevance Section: The portion of context directly related to a retrieved fact or result.
  • Retrieval Accuracy: The precision or correctness of results retrieved by a model.
  • Retrieval Method: The technique or algorithm used to fetch relevant data from a collection.
  • Retrieval Module: A component responsible for handling data retrieval in an application.
  • Reranker: An algorithm or model that reorders retrieval results to enhance relevance.

S

  • Secret Access: Restricted access for sensitive data, limited to specific authorized users.
  • Self-Contained Facts: Facts that are complete and independent, not requiring external context to be understood.
  • Similarity Score: A measurement indicating the degree of similarity between two items.
  • Similarity Search: A search method used to find items similar to a query based on similarity metrics.
  • Structured Data: Data organized in a defined format, like rows and columns in a table.
  • Server Cost: The cost associated with server hosting for applications or data.
  • Sparse Gradient Updates: Updates to model parameters that only involve a subset of gradients, often used for efficiency.

T

  • Token Generation: The process of creating tokens used for secure API access.
  • Top K: The number of top-ranking results retrieved, determined by relevance or similarity.
  • Transformer Layer Dropout: A regularization technique applied in transformer models where certain layers are randomly dropped during training to prevent overfitting.

U

  • Unstructured Data: Data without a predefined structure, such as text, images, or audio.
  • Unstructured Data with Metadata: Unstructured data accompanied by additional context, like tags or timestamps.

V

  • Verification Code: A code sent to verify user identity during account setup or login.
  • Version Control: The management of changes to documents, code, or data, tracking updates and preserving previous versions.
  • Variance Scaling Initialization: A method of initializing neural network weights to help prevent vanishing or exploding gradients during training.
  • Vector Quantization in Embeddings: A technique that reduces the dimensionality of embeddings by grouping similar vectors into a single representative vector, often used for efficient retrieval.