Glossary of Artificial Intelligence (AI) Terms

Classification of how independently an artificial intelligence agent can operate, ranging from Level 0 with no autonomy to Level 5 with full autonomy. Guidewire currently focuses on autonomy levels up to Level 2 (L2), which require human oversight and approval for any decisions or actions applied in core systems.

A multistep pipeline within the agent framework that chains together multiple agents, tools, and data transformations. Developers can use flows to orchestrate complex business logic with shared typed state management, maintaining conversation history and prior step outputs.

An approach to creating artificial intelligence assistants and agents that builds on retrieval-augmented generation by combining information retrieval with complex reasoning, multistep planning, and tool use. An assistant built with this approach can autonomously call tools to fetch data, analyze the retrieved information, and merge the results to execute specific tasks.

A goal-directed software application powered by large language models. It executes defined tasks through reasoning, prepares a plan, and autonomously directs its own processes and tool usage within guardrails to achieve its goal.

A reactive, software-embedded tool designed to support a human user by performing specific, context-aware tasks upon request. It operates under the human user’s permissions and runs only when prompted, providing recommendations or drafted content for human review rather than executing end-to-end business outcomes independently.

A Guidewire platform service providing access to artificial intelligence models, optical character recognition, and personal data redaction. It serves as the secure routing layer connecting applications and agents to large language models, including third-party services and a customer's own self-hosted or fine-tuned models.

A Guidewire RAG-as-a-service platform used with the Guidewire Agentic Framework that provides storage and retrieval to support retrieval-augmented generation. Agents and flows use it to securely access and search authoritative documents to ground their answers in factual data.

A largely deterministic pipeline that executes a sequence of tasks, leveraging an artificial intelligence model for one or more steps to provide probabilistic input to the next task.

A mechanism in large language models that enables the system to process and understand text by dynamically focusing on different parts of an input sequence. Instead of processing text sequentially, attention weighs the relative importance of each token, capturing nuances and dependencies across the text.

A general software program that executes predefined, rule-based tasks deterministically to deliver reliable and fast outcomes. It serves as a traditional, non-AI baseline to contrast against AI workflows and AI agents.

Automation executes predefined, rule-based tasks deterministically. An AI workflow is a largely deterministic pipeline that incorporates AI models for one or more steps. AI agents differ by using AI models to autonomously execute non-deterministic, adaptive tasks, dynamically planning and choosing tools.

The process of breaking down input text or documents into smaller, manageable units called chunks. This enhances the efficiency and accuracy of retrieval-augmented generation by allowing the model to focus on the most relevant text segments rather than processing entire documents at once.

In the Guidewire AI taxonomy, a competency is functionality that demonstrates subject matter expertise, understands how companies and systems work to get a job done, and has a clearly definable, measurable business outcome. A competency aggregates multiple AI and other capabilities and represents a meaningful transformation in how work gets done. Agents are designed to use competencies to complete tasks.

The information and signals that a model uses alongside the user’s prompt, such as prior messages, retrieved content, and system instructions. Context shapes how the model interprets a request and what output it produces.

The practice of selecting, structuring, and managing context so a model has the right information in the right form at the right time. Context engineering includes deciding what content, messages, and metadata to include or exclude, how that content is structured, and how it is tagged so the model can produce grounded, relevant responses.

An AI-powered coding assistant designed for developers working in the Guidewire ecosystem (customers, partners, and Guidewire teams) to enhance productivity and code quality. Developer Assistant understands the Guidewire development framework, including languages, libraries, cloud services, and best practices, and can assist with tasks such as explaining code, generating code, refactoring, and writing tests.

A numerical representation of text, also called vector embeddings, that captures the semantic meaning and context of words or sentences as vectors in a continuous space. This allows retrieval systems to match relevant facts with user queries based on underlying concepts rather than exact keyword matches.

Metrics used to measure the performance, correctness, and safety of an artificial intelligence agent. Online evaluations compute metrics such as latency and hallucination detection during live runs, while offline evaluations process batch metrics over datasets to validate agents before production deployment.

A prompt engineering technique where a large language model is provided with a few examples of a task to enhance its performance. By including these examples, the model better understands the desired output format and context.

A Guidewire serverless computing unit in which AI agents are deployed and executed on Guidewire Cloud Platform. Agents are exposed as functions with associated HTTP endpoints, scopes, and permissions.

A database that stores data as nodes (entities) and edges (relationships). Graph databases are optimized for traversing connections, making them useful for knowledge graphs, recommendation systems, and RAG patterns that rely on structured relationships.

Rules and constraints put in place to ensure artificial intelligence systems behave safely and accurately. They prevent hallucinations, block queries containing personal identifiable information, and restrict models to authorized boundaries and data sources.

An artificial intelligence agent development framework from Guidewire for building intelligent, tool-using agents and multistep workflows on Guidewire Cloud Platform. It wraps underlying AI services while providing abstractions for orchestration, tools, evaluations, state management, and deployment.

A scenario where a large language model generates text that is nonsensical, factually incorrect, or unfaithful to the provided source content. Guidewire mitigates this using retrieval-augmented generation and strict query guardrails to ensure responses are based exclusively on validated source documents.

In the Guidewire Agentic Framework, a handler defines what happens at a step in an agentic flow. It can be a handler function you write or a declarative agent handler specified in the step configuration.

A system design approach that integrates human oversight and intervention points within automated processes. It ensures users retain control by providing recommendations or drafted content for review and approval before actions are executed.

A retrieval method that blends semantic search with traditional keyword-based search. It leverages the precise matching of specific keywords alongside the contextual understanding of semantic search to deliver more accurate results.

The underlying goal or purpose behind a user’s query or action. Intent represents what the user is trying to achieve so the system can choose the right workflow, agent, or content.

The process of classifying a user's query to determine their underlying goal or intent before retrieving content. This helps the system route the query to the correct content retriever or identify if the query is out of scope.

An advanced artificial intelligence neural network trained on massive datasets to understand language and predict the next word in a sequence. It is used as the reasoning engine within assistants and agents to summarize, translate, and generate text based on prompts.

A reranking technique used in retrieval-augmented generation to reorder retrieved facts and create a more diverse set of information. By balancing pure relevance with diversity, MMR prevents the system from returning multiple text chunks containing highly similar information.

A server that implements the Model Context Protocol, exposing a set of tools, resources, or prompt templates that an agent can discover and invoke.

An open standard that defines how AI agents communicate with external tools, resources, and prompt templates.

The ability to monitor and track the internal state and operations of an artificial intelligence system. In the agent framework, this includes tracking threads, execution runs, evaluation metrics, and structured logging to simplify debugging.

The process of crafting, refining, and optimizing the natural language instructions given to a large language model to guide its behavior. Effective prompts provide context, constraints, and specific output formats to achieve accurate and predictable results.

ProNavigator is an AI-powered knowledge management platform built specifically for the insurance industry. It gives insurance professionals — such as agents, brokers, underwriters, and claims adjusters — instant access to up-to-date information such as policy wordings, underwriting guidelines, and insurer documents, all from one centralized platform. The goal is to help teams work faster, reduce errors, and serve customers more efficiently.

A second-stage process in a retrieval-augmented generation pipeline that reassesses and reorders an initially retrieved set of documents to enhance relevance. It applies a sophisticated algorithm to ensure only the highest-quality information is passed to the generative model.

A technique that provides a large language model with relevant contextual information retrieved from an external knowledge base. By grounding the response in factual, authorized data external to its training dataset, this process improves accuracy and reduces hallucinations and serves as the base pattern on which agentic RAG builds.

A search method that considers the user's intent and the contextual meaning of terms using vector embeddings, rather than relying solely on exact keyword matches. This allows the system to understand underlying concepts and deliver more precise results.

A discrete, goal-oriented capability that an agent can autonomously select and execute. A skill captures the agent’s procedural knowledge about how to carry out a task and is what the agent uses to translate intent into actions.

A specialized agent spawned by a parent (orchestrator) to handle a subtask in isolation. A subagent operates in a dedicated context window, preventing the parent’s history from being flooded with low-level logs or intermediate data.

A unit of work an agent is trying to complete. A task represents a specific goal or outcome, and agents use skills and tools to carry out the steps needed to complete it.

The process of breaking down a text string into smaller, manageable units called tokens, such as single characters, individual words, or subwords. This is a foundational step required for large language models to process language.

A primitive, external interface that allows an agent to interact with systems and data sources beyond its internal model. A tool is a passive object with no reasoning of its own; it simply executes a specific function when called with the correct inputs.

A specialized database, often referred to as a vector store, optimized for storing and managing data keyed by high-dimensional vectors. It provides efficient query capabilities, such as similarity search, which is essential for quickly retrieving relevant document chunks during retrieval-augmented generation.