Agentic glossary¶

This document contains all terms from the DataRobot glossary that pertain to agents or agentic workflows.

A¶

Agent¶

An AI-powered component within DataRobot designed to execute complex, multi-step tasks autonomously. An agent can be configured with specific goals, LLMs, and a set of tools, allowing it to perform actions like orchestrating a data preparation workflow, running a modeling experiment, or generating an analysis without direct human intervention. Agents exhibit autonomous behavior, can reason about their environment, make decisions, and adapt their strategies based on feedback. Multiple agents can be combined in an agentic workflow to solve more sophisticated business problems through collaboration and coordination.

Agent-based modeling¶

Computational modeling approaches that simulate complex systems by modeling individual agents and their interactions. Agent-based modeling enables the study of emergent behaviors and system-level properties that arise from individual agent behaviors. In DataRobot's platform, agent-based modeling capabilities allow users to simulate business processes, test agent strategies, and understand how different agent configurations affect overall system performance.

Agentic AI¶

A paradigm of artificial intelligence where AI systems are designed to act as autonomous agents that can perceive their environment, reason about goals, plan actions, and execute tasks with minimal human oversight. Agentic AI systems are characterized by their ability to make independent decisions, learn from experience, and adapt their behavior to achieve objectives. In DataRobot's platform, agentic AI enables sophisticated automation of complex data science workflows, allowing AI systems to handle end-to-end processes from data preparation to model deployment and monitoring.

Agentic workflow¶

Systems that leverage AI agents to perform tasks and make decisions within a workflow, often with minimal human intervention. Agentic workflows can be built in a local IDE using DataRobot templates and a CLI and managed with real-time LLM intervention and moderation with out-of-the-box and custom guards, including integration with NVIDIA's NeMo for content safety and topical rails in the UI or with code.

Agent Framework (AF) components¶

Agent Framework (AF) components provide modular building blocks for constructing sophisticated AI agents. AF components include reasoning engines, memory systems, action planners, and communication modules that can be combined to create custom agent architectures. In DataRobot's platform, AF components enable rapid development of specialized agents with specific capabilities while maintaining consistency and interoperability across different agent implementations.

Agent-to-Agent (A2A)¶

Agent-to-Agent (A2A) refers to communication protocols and frameworks that enable direct interaction and coordination between AI agents. A2A systems facilitate information sharing, task delegation, and collaborative problem-solving among multiple agents. In DataRobot's agentic workflows, A2A capabilities enable agents to work together seamlessly, share context and knowledge, and coordinate complex multi-agent operations while maintaining security and governance controls.

Alignment¶

The critical process of steering an AI model's outputs and behavior to conform to an organization's specific ethical guidelines, safety requirements, and business objectives. In DataRobot, alignment is practically applied through features like guardrails, custom system prompts, and content moderation policies. This practice helps to mitigate risks from biased, unsafe, or off-topic model responses, ensuring the AI remains a trustworthy and reliable tool for the enterprise.

Apps¶

See No-Code AI Apps.

Autonomy¶

The ability of an AI agent to operate independently and make decisions without constant human oversight. Autonomous agents can plan, execute, and adapt their behavior based on changing conditions and feedback. In DataRobot's agentic workflows, autonomous capabilities are balanced with human oversight through guardrails and monitoring to ensure safe and effective operation. Autonomy enables agents to handle complex, multi-step processes while maintaining alignment with business objectives and safety requirements.

C¶

Chain-of-thought¶

A prompting technique that encourages language models to break down complex problems into step-by-step reasoning processes. In DataRobot's agentic workflows, chain-of-thought prompting enhances agent reasoning capabilities by requiring explicit intermediate steps in decision-making, leading to more transparent and reliable outcomes. This technique improves problem-solving accuracy and enables better debugging and validation of agent behavior in multi-step tasks.

Chatting¶

Sending prompts (and as a result, LLM payloads) to LLM endpoints based on a single LLM blueprint and receiving a response from the LLM. In this case, context from previous prompts/responses is sent along with the payload.

Chunking¶

The action of taking a body of unstructured text and breaking it up into smaller pieces of unstructured text (tokens).

Citation¶

The chunks of text from the vector database used during the generation of LLM responses.

CLI¶

Command Line Interface (CLI) tools that enable programmatic interaction with DataRobot's agentic workflows and platform services. CLI tools provide scriptable access to agent configuration, workflow execution, and platform management functions. In DataRobot's agentic ecosystem, CLI tools support automation of agent deployment, monitoring, and maintenance tasks, enabling integration with CI/CD pipelines and automated workflows.

Cognitive architecture¶

The underlying structural framework that defines how AI agents process information, make decisions, and interact with their environment. Cognitive architectures specify the components, processes, and relationships that enable intelligent behavior in agents. In DataRobot's agentic workflows, cognitive architectures provide the foundation for agent reasoning, memory management, learning, and decision-making capabilities, enabling sophisticated autonomous behavior.

Connected vector database¶

An external vector database accessed via a direct connection to a supported provider for vector database creation. The data source is stored locally in the Data Registry, configuration settings are applied, and the created vector database is written back to the provider. Connected vector databases maintain real-time synchronization with the platform and provide seamless access to embeddings and text chunks for grounding LLM responses.

Context window¶

The limited amount of information, measured in tokens, that a large language model can hold in its active memory for a single chat conversation turn. This 'memory' includes the user's prompt, any recent conversation history provided, and data retrieved via Retrieval Augmented Generation (RAG). The size of the context window is a critical parameter in an LLM blueprint, as it dictates the model's ability to handle long documents or maintain coherence over extended dialogues; any information outside this window is not considered when generating the next response.

Conversation memory¶

The ability of an AI system to remember and reference previous interactions within a conversation session (meaning that the session contains one or more chat conversation turns). Conversation memory enables contextual continuity, allowing the AI to maintain awareness of earlier exchanges and build upon previous responses. In DataRobot's chat interfaces, conversation memory helps maintain coherent, contextually relevant dialogues.

D¶

Deploying (from a playground)¶

LLM blueprints and all their associated settings are registered in Registry and can be deployed with DataRobot's production suite of products.

Directed acyclic graph (DAG)¶

A mathematical structure used to represent workflows where nodes represent tasks or operations and edges represent dependencies between them. In AI workflows, DAGs ensure that tasks are executed in the correct order without circular dependencies, enabling efficient orchestration of complex multi-step processes like data preprocessing, model training, and deployment pipelines.

E¶

Embedding¶

A numerical (vector) representation of text, or a collection of numerical representations of text. The action of generating embeddings means taking a chunk of unstructured text and using a text embedding model to convert the text to a numerical representation. The chunk is the input to the embedding model and the embedding is the "prediction" or output of the model.

Episodic memory¶

Memory systems that store specific experiences, events, and contextual information about past interactions and situations. Episodic memory enables AI agents to recall specific instances, learn from particular experiences, and apply contextual knowledge to similar situations. In DataRobot's agentic workflows, episodic memory allows agents to remember specific user interactions, successful task executions, and contextual details that inform future decision-making.

F¶

Few-shot learning¶

A capability of a model to learn to perform a task from a small number of examples provided in the prompt.

Few-shot prompting¶

A technique where a few examples are provided in the prompt (either in an input or system prompt) to guide the model's behavior and improve its performance on specific tasks. Few-shot prompting helps models understand the desired output format and style without requiring fine-tuning, making it useful for quick adaptation to new tasks or domains.

Fine-tuning¶

The process of adapting pre-trained foundation models to specific tasks or domains by continuing training on targeted datasets. In DataRobot's platform, fine-tuning enables users to customize large language models for particular use cases, improving performance on domain-specific tasks while preserving general capabilities. Unlike prompt engineering which works with existing model weights, fine-tuning modifies the model's internal parameters to create specialized versions optimized for particular applications, industries, or data types.

Foundation model¶

A powerful, large-scale AI model, like GPT or Claude, that provides broad, general-purpose capabilities learned from massive datasets. In the DataRobot platform, these models act as the core component or 'foundation' of an LLM blueprint. Rather than being a ready-made solution, a foundation model is the versatile starting point that can be customized for specific business needs through techniques like prompting, RAG, or fine-tuning.

Function calling¶

The capability of large language models to invoke external functions, tools, or APIs based on user requests and conversation context. In DataRobot's agentic workflows, function calling enables agents to perform actions beyond text generation, such as data retrieval, mathematical computations, API interactions, and system operations. This allows agents to execute complex tasks, integrate with enterprise systems, and provide dynamic responses based on real-time information. Function calling transforms conversational AI into actionable systems that can manipulate data and interact with external services.

G¶

Generative AI (GenAI)¶

A type of artificial intelligence that generates new content based on learned patterns from training data. In DataRobot's platform, GenAI capabilities include text generation, content creation, and intelligent responses through LLM blueprints. Unlike traditional predictive models that analyze existing data, GenAI creates novel outputs through prompting and can be integrated into DataRobot workflows for content generation, analysis, and automated decision-making processes.

Guardrails¶

Safety mechanisms that prevent AI systems from generating harmful or inappropriate content. Guardrails include content filtering, output validation, and behavioral constraints that ensure AI responses align with safety guidelines and organizational policies. In DataRobot, guardrails can be configured and help maintain responsible AI practices and prevent the generation of unsafe or unethical content.

Grounding¶

The process of ensuring that language model responses are based on specific, verifiable data sources rather than relying solely on training data. In DataRobot's platform, grounding is achieved through Retrieval Augmented Generation (RAG) workflows that connect LLMs to vector databases containing relevant documents, knowledge bases, or enterprise data. This technique improves response accuracy, reduces hallucinations, and ensures that AI outputs are contextualized with current, relevant information from trusted sources.

H¶

Hallucination¶

When a language model generates information that is plausible-sounding but factually incorrect or not grounded in the provided data.

High code¶

A development approaches that emphasizes custom programming and fine-grained control over application behavior. High-code solutions provide maximum flexibility and customization capabilities for complex requirements. In DataRobot's agentic workflows, high-code capabilities enable advanced users to create highly specialized agents with custom logic, integrate with complex enterprise systems, and implement sophisticated decision-making algorithms.

Human in the loop (HILT)¶

Integration patterns that incorporate human oversight, validation, and intervention into AI agent workflows. Human-in-the-loop systems enable humans to review agent decisions, provide feedback, correct errors, and guide agent behavior at critical decision points. In DataRobot's agentic workflows, human-in-the-loop capabilities ensure quality control, enable learning from human expertise, and maintain human authority over sensitive or high-stakes decisions.

I¶

In-context learning¶

The ability of LLMs to learn from examples provided in the prompt without requiring fine-tuning. In-context learning allows models to adapt their behavior based on the context and examples given in the current conversation, enabling them to perform new tasks or follow specific instructions without additional training.

Instruction tuning¶

Training LLMs to follow specific instructions or commands by fine-tuning them on instruction-response pairs. Instruction tuning improves a model's ability to understand and execute user requests, making it more useful for practical applications where following directions is important.

K¶

Knowledge cutoff¶

The date after which an LLM's training data ends, limiting its knowledge of historical events, information, and developments that occurred after that point. Knowledge cutoff dates are important for understanding the temporal scope of a model's information and determining when additional context or real-time data sources may be needed.

L¶

Large language model (LLM)¶

A deep learning model trained on extensive text datasets that can understand, generate, and process human language. In DataRobot's platform, LLMs form the core of LLM blueprints and can be configured with various settings, system prompts, and vector databases to create customized AI applications. These models enable DataRobot users to build intelligent chatbots, content generators, and analysis tools that can understand context and provide relevant responses.

LLM blueprint¶

The saved blueprint, available to be used for deployment. LLM blueprints represent the full context for what is needed to generate a response from an LLM; the resulting output can be compared within the playground. This information is captured in the LLM blueprint settings.

LLM blueprint components¶

The entities that make up the LLM blueprint settings, this refers to the vector database, embedding model user to generate the vector database, LLM settings, system prompt, etc. These components can either be offered natively within DataRobot or can be brought in from external sources.

LLM blueprint settings¶

The parameters sent to the LLM to generate a response (in conjunction with the user-entered prompt). They include a single LLM, LLM settings, optionally a system prompt, and optionally a vector database. If no vector database is assigned, then the LLM uses its learnings from training to generate a response. LLM blueprint settings are configurable so that you can experiment with different configurations.

LLM gateway¶

A centralized service in DataRobot that manages access to multiple large language models from external providers with support for unified authentication, rate limiting, and request routing. The LLM gateway enables organizations to standardize their interactions with various LLM providers while maintaining security, monitoring, and cost controls across all model usage.

LLM payload¶

The bundle of contents sent to the LLM endpoint to generate a response. This includes the user prompt, LLM settings, system prompt, and information retrieved from the vector database.

LLM responses¶

Generated text from the LLM based on the payload sent to an LLM endpoint.

LLM settings¶

Parameters that define how an LLM intakes a user prompt and generates a response. They can be adjusted within the LLM blueprint to alter the response. These parameters are currently represented by the "Temperature", "Token selection probability cutoff (Top P)", and "Max completion tokens" settings.

Low code¶

A development approach that minimizes the amount of manual coding required to build applications and workflows. Low-code platforms provide visual interfaces, drag-and-drop components, and pre-built templates that enable rapid development. In DataRobot's agentic workflows, low-code capabilities allow users to create sophisticated AI agents and workflows through configuration interfaces rather than extensive programming, making agentic AI accessible to non-technical users.

M¶

Multi-agent flow¶

A workflow pattern where multiple AI agents collaborate to solve complex problems by dividing tasks among specialized agents. Each agent has specific capabilities and responsibilities, and they communicate and coordinate to achieve the overall objective. Multi-agent flows enable more sophisticated problem-solving by leveraging the strengths of different specialized agents. See also Agentic workflow.

Model Context Protocol (MCP) server¶

A Model Context Protocol (MCP) server provides standardized interfaces for AI agents to interact with external systems and data sources. MCP servers enable secure, controlled access to tools, databases, APIs, and other resources that agents need to accomplish their tasks. In DataRobot's agentic workflows, MCP servers facilitate seamless integration between agents and enterprise systems while maintaining security and governance controls.

Model alignment¶

Techniques to ensure AI models behave according to human values and intentions. Model alignment involves training and fine-tuning processes that help models produce outputs that are helpful, honest, and harmless, reducing risks of harmful or unintended behaviors in production environments.

Moderation¶

The process of monitoring and filtering model outputs to ensure they comply with safety, ethical, and policy guidelines.

N¶

NAT¶

Neural Architecture Transfer (NAT) enables efficient transfer of learned representations and architectures between different AI models and tasks. NAT techniques allow agents to leverage pre-trained components and adapt them for specific use cases without full retraining. In DataRobot's agentic workflows, NAT capabilities enable rapid deployment of specialized agents by transferring knowledge from general-purpose models to domain-specific applications.

NIM¶

NVIDIA Inference Microservice (NIM) is a containerized AI model that provides optimized, high-performance inference with low latency and efficient resource utilization. In DataRobot's platform, a NIM can be integrated into an agentic workflow to provide specialized AI capabilities, enabling agents to leverage state-of-the-art models for specific tasks while maintaining optimal performance and scalability.

O¶

One-shot learning¶

A capability of a model to learn to perform a task from only a single example.

Orchestration¶

The coordination of multiple AI components, tools, and workflows to achieve complex objectives. Orchestration involves managing the flow of data and control between different AI services, ensuring proper sequencing, error handling, and resource allocation. In DataRobot, orchestration enables the creation of sophisticated multi-step AI workflows that combine various capabilities and tools.

P¶

Parameter efficient fine-tuning (PEFT)¶

Methods to fine-tune large models using fewer parameters than full fine-tuning. PEFT techniques, such as LoRA (Low-Rank Adaptation) and adapter layers, allow for efficient model customization while maintaining most of the original model's performance and reducing computational requirements.

Playground¶

The place where you create and interact with LLM blueprints (LLMs and their associated settings), comparing the response of each to help determine which to use in production. Many LLM blueprints can live within a playground. A playground is an asset of a Use Case; multiple playgrounds can exist in a single Use Case.

Playground compare¶

The place to add LLM blueprints to the playground for comparison, submit prompts to these LLM blueprints, and evaluate the rendered responses. With RAG, a single prompt is sent to an LLM to generate a single response, without referencing previous prompts. This allows users to compare responses from multiple LLM blueprints.

Prompt¶

The input entered during chatting used to generate the LLM response.

Prompt engineering¶

The practice of designing and refining input prompts to guide a language model toward producing desired outputs.

Prompt injection¶

A security vulnerability where malicious prompts can override system instructions or safety measures. Prompt injection attacks attempt to manipulate AI systems into generating inappropriate content or performing unintended actions by crafting inputs that bypass the model's intended constraints and guidelines.

Prompt template¶

See system prompt.

Pulumi¶

Infrastructure as Code (IaC) platform that enables developers to define and manage cloud infrastructure using familiar programming languages. Pulumi supports multiple cloud providers and provides a unified approach to infrastructure management. In DataRobot's agentic workflows, Pulumi enables automated provisioning and management of infrastructure resources needed for agent deployment, scaling, and monitoring across different environments.

R¶

Reinforcement learning from human feedback (RLHF)¶

A training method that uses human feedback to improve model behavior. RLHF involves collecting human preferences on model outputs and using reinforcement learning techniques to fine-tune the model to produce responses that align with human values and preferences, improving safety and usefulness.

ReAct¶

A Reasoning and Acting (ReAct) framework combines reasoning capabilities with action execution in AI agents. ReAct enables agents to think through problems step-by-step, plan actions, execute them, and observe results to inform subsequent reasoning. In DataRobot's agentic workflows, ReAct capabilities allow agents to perform complex problem-solving by iteratively reasoning about situations, taking actions, and learning from outcomes to achieve their goals.

Retrieval¶

The process of finding relevant information from a knowledge base or database. In the context of RAG workflows, retrieval involves searching through vector databases or other knowledge sources to find the most relevant content that can be used to ground and inform AI responses, improving accuracy and reducing hallucination.

Retrieval Augmented Generation (RAG)¶

The process of sending a payload to an LLM that contains the prompt, system prompt, LLM settings, vector database (or subset of vector database), and the LLM returning corresponding text based on this payload. It includes the process of retrieving relevant information from a vector database and sending that along with the prompt, system prompt, and LLM settings to the LLM endpoint to generate a response grounded in the data in the vector database. This operation may optionally also incorporate orchestration to execute a chain of multiple prompts.

Retrieval Augmented Generation (RAG) workflow¶

An AI system that runs RAG, which includes data preparation, vector database creation, LLM configuration, and response generation. RAG workflows typically involve steps such as document chunking, embedding generation, similarity search, and context-aware response generation, all orchestrated to provide accurate, grounded responses to user queries. See also Retrieval Augmented Generation (RAG).

S¶

Semantic memory¶

Memory systems that store general knowledge, facts, concepts, and relationships that are not tied to specific experiences. Semantic memory enables AI agents to maintain domain knowledge, understand concepts, and apply general principles to new situations. In DataRobot's agentic workflows, semantic memory allows agents to maintain knowledge about business processes, domain expertise, and general problem-solving strategies.

Semantic search¶

Search method that finds content based on meaning rather than exact keyword matches. Semantic search uses vector embeddings to understand the intent and context of queries, enabling more accurate and relevant results even when the exact words don't match. This approach is particularly useful in RAG systems for finding the most relevant information to ground AI responses.

Short-term memory¶

Temporary storage systems that AI agents use to maintain context and information during active task execution. Short-term memory enables agents to remember recent interactions, maintain conversation context, and track progress on current tasks. In DataRobot's agentic workflows, short-term memory allows agents to maintain coherence across multi-step processes and provides continuity in user interactions.

Long-term memory¶

Persistent storage systems that AI agents use to retain knowledge, experiences, and learned patterns across multiple sessions and tasks. Long-term memory enables agents to build upon previous experiences, maintain learned behaviors, and accumulate domain knowledge over time. In DataRobot's agentic workflows, long-term memory allows agents to improve performance through experience and maintain consistency across different use cases.

Streaming¶

Real-time generation of text where output is displayed as it's being generated. Streaming provides immediate feedback to users by showing AI responses as they are produced, rather than waiting for the complete response. This approach improves user experience by reducing perceived latency and allowing users to see progress in real-time.

Single agent flow¶

A workflow pattern where a single AI agent handles all aspects of a task from start to finish. The agent receives input, processes it through its capabilities, and produces output without requiring coordination with other agents. Single agent flows are suitable for straightforward tasks that can be completed by one specialized agent.

Sidecar model¶

A structural component that supports the LLM that is serving back responses. It helps to make determinations about whether a prompt was toxic, an injection attack, etc. With DataRobot, it uses hosted custom metrics to accomplish the monitoring.

Stop sequence¶

A specific token or set of tokens that signals a language model to stop generating further output.

Syftr¶

A specialized agent framework component that provides secure, privacy-preserving data processing capabilities for AI agents. Syftr enables agents to work with sensitive data while maintaining confidentiality and compliance with privacy regulations. In DataRobot's agentic workflows, Syftr components allow agents to process encrypted or anonymized data, perform federated learning, and maintain data privacy throughout the agent lifecycle.

System prompt¶

The system prompt, an optional field, is a "universal" prompt prepended to all individual prompts. It instructs and formats the LLM response. The system prompt can impact the structure, tone, format, and content that is created during the generation of the response.

T¶

Temperature¶

A parameter that controls the creativity and randomness of LLM responses. Lower temperature values (0.1-0.3) produce more focused, consistent outputs suitable for factual responses, while higher values (0.7-1.0) generate more creative and diverse content. DataRobot's playground interface allows you to experiment with different temperature values in LLM blueprint settings to find the optimal balance for your specific use case.

Template¶

Pre-configured frameworks or structures that provide a starting point for creating agentic workflows, applications, or configurations. Templates in DataRobot include predefined agent configurations, workflow patterns, and code structures that accelerate development and ensure best practices. Templates can include agent goals, tool configurations, guardrails, and integration patterns, allowing users to quickly deploy sophisticated agentic systems without starting from scratch.

Token¶

The smallest unit of text that LLMs process when parsing prompts/generating responses. In DataRobot's platform, tokens are used to measure input/output size of chats and calculate usage costs for LLM operations. When you send prompts to LLM blueprints, the system tokenizes your text and tracks consumption for billing and performance monitoring. Token usage is displayed in DataRobot's playground and deployment interfaces to help you optimize costs and stay within platform limits.

Tokenization¶

The process of breaking text into smaller units called tokens, which can be words, subwords, or characters, for processing by a language model.

Token usage¶

The number of tokens consumed by an LLM for input and output, often used for billing and cost management. Token usage is a key metric for understanding the computational cost of AI operations, as most LLM providers charge based on the number of tokens processed. Monitoring token usage helps optimize costs and resource allocation in AI applications.

Tool¶

A software component or service that provides specific functionality to AI agents or workflows. Tools can perform various tasks such as data retrieval, computation, API calls, or specialized processing. In DataRobot's agentic workflows, tools are modular components that agents can invoke to extend their capabilities and perform complex operations beyond their core functionality.

Toolkit¶

A collection of tools, utilities, and resources designed to support the development and deployment of agentic AI systems. Toolkits provide standardized interfaces, common functionality, and best practices for building AI agents. In DataRobot's platform, toolkits include pre-built tools for data processing, model training, API integration, and workflow orchestration, enabling rapid development of sophisticated agentic applications.

Top-k¶

A decoding parameter that limits the model's next-token choices to the k most likely options, sampling from only those candidates to generate more focused or creative responses.

Top-p (nucleus sampling)¶

A decoding parameter that limits the model's next-token choices to the smallest set whose cumulative probability exceeds a threshold p, allowing for dynamic selection of likely tokens.

Toxicity¶

The presence of harmful, offensive, or inappropriate language in model outputs, which safety and moderation systems aim to detect and prevent.

U¶

Unstructured text¶

Text that cannot fit cleanly into a table. The most typical example is large blocks of text typically in some kind of document or form.

V¶

Vector database¶

A specialized database that stores text chunks alongside their numerical representations (embeddings) for efficient similarity search. In DataRobot's platform, vector databases enable RAG operations by allowing LLM blueprints to retrieve relevant information from large document collections. When you upload documents to DataRobot, the system automatically chunks the text, generates embeddings, and stores them in a vector database that can be connected to LLM blueprints for grounded, accurate responses based on your specific content.

Z¶

Zero-shot learning¶

A capability of a model to perform a task without having seen any examples of that task during training, relying on generalization from related knowledge.