> ## Documentation Index > Fetch the complete documentation index at: https://www.adaline.ai/docs/llms.txt > Use this file to discover all available pages before exploring further. # Glossary > Definitions of common terms and concepts in Adaline Use the built-in search (`Cmd+K` or `Ctrl+K`) to quickly find terms, or browse alphabetically using the tabs below. A local SDK that provides a unified interface for calling 200+ LLMs with built-in batching, retries, caching, callbacks, and OpenTelemetry support. Also referred to as the "Super SDK" for its cross-provider compatibility. The Adaline log proxy that intercepts LLM API calls and sends them to Monitor without modifying application code. A production-derived evaluator created from logs, sessions, Behaviors, feedback, or representative examples. Auto Generated Evaluators help Improve cycles score candidates against issues found in real traffic. A recurring pattern detected from project evidence, such as a user intent, assistant response pattern, tool behavior, issue, drift signal, healthy workflow, or agent trajectory pattern. An automated evaluation process that runs on live data in Monitor. Samples incoming requests and scores them using configured evaluators to track quality over time. A collection of test cases stored in Adaline used for evaluations. Each dataset consists of rows and columns that can map to prompt variables, expected outputs, labels, evaluator inputs, or metadata. A semantic search workflow for finding relevant logs by meaning rather than exact tags, filters, or keywords. Deep Search is used when you know the behavior you want to investigate but not the exact trace metadata. A versioned snapshot of a prompt configuration that has been published to a specific environment. Deployments are accessible via API or SDK and can be rolled back to previous versions. A named runtime lane, such as staging or production, that points an application to a deployed prompt snapshot. Environments let teams release, compare, and roll back prompt versions safely. A dataset column that fetches its value from an external API or another prompt at runtime, rather than storing static values. Enables live data in evaluations. The process of running a prompt against a dataset and scoring the responses using one or more evaluators. Evaluations produce reports with pass/fail rates and detailed metrics. A configured metric or judge used to assess prompt performance during evaluations. Types include LLM-as-a-Judge, JavaScript, Cost, Latency, Text Matcher, and more. A reviewed workflow that turns production evidence into a proposed prompt change. An Improve cycle can use Behaviors, logs, sessions, evaluators, datasets, and prompt optimization before pausing for human review. An evaluation pattern where an LLM scores the output of another LLM based on configurable criteria. Enables subjective quality assessment at scale. The Adaline surface for analyzing production logs, traces, spans, charts, cost, latency, token usage, feedback, and continuous evaluation signals. A segment inside a session or trajectory, such as planning, retrieval, editing, testing, recovery, or final answer. Phases help reviewers understand where an agent journey changed direction or failed. The interactive testing environment in Adaline's prompt editor. Allows running prompts with different inputs and comparing outputs side-by-side. A container in Adaline that holds prompts, datasets, evaluations, logs, Behaviors, Improve cycles, and deployments. A project can be considered equivalent to an AI agent, application, or workflow. A configured instruction template in Adaline consisting of messages, model settings, variables, and optional tools. Prompts are versioned and can be deployed to environments. A proposed prompt snapshot generated during an Improve cycle. Reviewers compare the candidate against the current prompt using diffs, evaluations, examples, and runtime tradeoffs before approving or rejecting it. The technique of connecting multiple prompts where the output of one becomes the input of another. Enables complex workflows with sequential LLM calls. A saved snapshot of a prompt's configuration at a specific point in time. Versions are auto-incremented and can be deployed or compared in evaluations. An LLM service provider (e.g., OpenAI, Anthropic, Google, Azure). Each provider offers different models with varying capabilities and pricing. Dataset rows, evaluators, and review evidence that preserve known failures or important healthy behavior so future prompt changes can be tested before release. A model configuration that constrains output structure (e.g., JSON mode, JSON schema). Ensures consistent, parseable responses from the LLM. A summarized journey grouped by a stable session ID. Sessions help Adaline understand multi-turn conversations, multi-step workflows, and agent tasks that cannot be judged from one trace alone. A single operation within a trace representing one LLM call or logical step. Spans have start/end times, attributes, and can be nested. A set of generated cases that expands evaluation coverage around production evidence, Behaviors, or known failure modes. Synthetic cases should be reviewed before they become release-grade coverage. A function that an LLM can call during generation. Tools have schemas defining their parameters and are used for retrieval, calculations, and external actions. A structured request from the LLM to execute a specific tool with provided arguments. Tool calls must be processed by the application and results returned to the LLM. A complete record of a user request flow containing multiple spans. Traces enable end-to-end visibility of LLM operations in Monitor. An agent-run journey built from sessions, phases, turns, and source spans. Trajectories are especially useful for multi-step and coding-agent workflows. A diagnosis attached to a Behavior, usually with severity, hypothesis, suggested fix, and links to representative evidence. A reconstructed step inside a session or phase, such as a user message, assistant reasoning step, tool call, tool result, or assistant reply. Turns can link back to source spans. A placeholder in prompt templates (e.g., `{{user_input}}`) that is replaced with actual values at runtime. Variables are mapped to dataset columns in evaluations. An HTTP callback triggered by Adaline events such as deployments or evaluation completions. Enables integration with external systems and CI/CD pipelines. The top-level organizational unit in Adaline containing projects, members, and API keys. Workspaces define billing boundaries and access control. A local SDK that provides a unified interface for calling 200+ LLMs with built-in batching, retries, caching, callbacks, and OpenTelemetry support. Also referred to as the "Super SDK" for its cross-provider compatibility. The Adaline log proxy that intercepts LLM API calls and sends them to Monitor without modifying application code. A production-derived evaluator created from logs, sessions, Behaviors, feedback, or representative examples. Auto Generated Evaluators help Improve cycles score candidates against issues found in real traffic. A recurring pattern detected from project evidence, such as a user intent, assistant response pattern, tool behavior, issue, drift signal, healthy workflow, or agent trajectory pattern. An automated evaluation process that runs on live data in Monitor. Samples incoming requests and scores them using configured evaluators to track quality over time. A collection of test cases stored in Adaline used for evaluations. Each dataset consists of rows and columns that can map to prompt variables, expected outputs, labels, evaluator inputs, or metadata. A semantic search workflow for finding relevant logs by meaning rather than exact tags, filters, or keywords. Deep Search is used when you know the behavior you want to investigate but not the exact trace metadata. A versioned snapshot of a prompt configuration that has been published to a specific environment. Deployments are accessible via API or SDK and can be rolled back to previous versions. A named runtime lane, such as staging or production, that points an application to a deployed prompt snapshot. Environments let teams release, compare, and roll back prompt versions safely. A dataset column that fetches its value from an external API or another prompt at runtime, rather than storing static values. Enables live data in evaluations. The process of running a prompt against a dataset and scoring the responses using one or more evaluators. Evaluations produce reports with pass/fail rates and detailed metrics. A configured metric or judge used to assess prompt performance during evaluations. Types include LLM-as-a-Judge, JavaScript, Cost, Latency, Text Matcher, and more. A reviewed workflow that turns production evidence into a proposed prompt change. An Improve cycle can use Behaviors, logs, sessions, evaluators, datasets, and prompt optimization before pausing for human review. An evaluation pattern where an LLM scores the output of another LLM based on configurable criteria. Enables subjective quality assessment at scale. The Adaline surface for analyzing production logs, traces, spans, charts, cost, latency, token usage, feedback, and continuous evaluation signals. A segment inside a session or trajectory, such as planning, retrieval, editing, testing, recovery, or final answer. Phases help reviewers understand where an agent journey changed direction or failed. The interactive testing environment in Adaline's prompt editor. Allows running prompts with different inputs and comparing outputs side-by-side. A container in Adaline that holds prompts, datasets, evaluations, logs, Behaviors, Improve cycles, and deployments. A project can be considered equivalent to an AI agent, application, or workflow. A configured instruction template in Adaline consisting of messages, model settings, variables, and optional tools. Prompts are versioned and can be deployed to environments. A proposed prompt snapshot generated during an Improve cycle. Reviewers compare the candidate against the current prompt using diffs, evaluations, examples, and runtime tradeoffs before approving or rejecting it. The technique of connecting multiple prompts where the output of one becomes the input of another. Enables complex workflows with sequential LLM calls. A saved snapshot of a prompt's configuration at a specific point in time. Versions are auto-incremented and can be deployed or compared in evaluations. An LLM service provider (e.g., OpenAI, Anthropic, Google, Azure). Each provider offers different models with varying capabilities and pricing. Dataset rows, evaluators, and review evidence that preserve known failures or important healthy behavior so future prompt changes can be tested before release. A model configuration that constrains output structure (e.g., JSON mode, JSON schema). Ensures consistent, parseable responses from the LLM. A summarized journey grouped by a stable session ID. Sessions help Adaline understand multi-turn conversations, multi-step workflows, and agent tasks that cannot be judged from one trace alone. A single operation within a trace representing one LLM call or logical step. Spans have start/end times, attributes, and can be nested. A set of generated cases that expands evaluation coverage around production evidence, Behaviors, or known failure modes. Synthetic cases should be reviewed before they become release-grade coverage. A function that an LLM can call during generation. Tools have schemas defining their parameters and are used for retrieval, calculations, and external actions. A structured request from the LLM to execute a specific tool with provided arguments. Tool calls must be processed by the application and results returned to the LLM. A complete record of a user request flow containing multiple spans. Traces enable end-to-end visibility of LLM operations in Monitor. An agent-run journey built from sessions, phases, turns, and source spans. Trajectories are especially useful for multi-step and coding-agent workflows. A diagnosis attached to a Behavior, usually with severity, hypothesis, suggested fix, and links to representative evidence. A reconstructed step inside a session or phase, such as a user message, assistant reasoning step, tool call, tool result, or assistant reply. Turns can link back to source spans. A placeholder in prompt templates (e.g., `{{user_input}}`) that is replaced with actual values at runtime. Variables are mapped to dataset columns in evaluations. An HTTP callback triggered by Adaline events such as deployments or evaluation completions. Enables integration with external systems and CI/CD pipelines. The top-level organizational unit in Adaline containing projects, members, and API keys. Workspaces define billing boundaries and access control.