Supported Use Cases

Deepchecks works with any LLM-powered application. Each application is composed of one or more sub-tasks, and each sub-task maps to an Interaction Type - a named configuration that determines which built-in properties are enabled, how auto-annotation rules are set up, and how interactions are displayed in the UI.

You can use one interaction type per application (for simple pipelines) or combine several (for complex agent workflows where retrieval, LLM calls, and tool use each need independent evaluation).

Not sure which type to choose? Start with the closest match - you can always change it later. Deepchecks' team is also happy to advise.

Classic Interaction Types

These types cover single-step or simple multi-step LLM tasks.

Interaction Type	Use case	Key properties
Q&A	Question answering and RAG pipelines	Grounded in Context, Retrieval Relevance, Completeness, Avoided Answer
Summarization	Condensing documents or transcripts	Coverage, Grounded in Context, Conciseness
Generation	Creative or structured content generation	Instruction Fulfillment, Fluency, Toxicity
Classification	Multi-class or multi-label LLM classification	Sentiment, Fluency, expected output comparison
Feature Extraction	Extracting structured data (e.g., JSON) from free text	Extraction Groundedness, Extraction Coverage, Structural Validity
Chat	Multi-turn conversational assistants	Instruction Fulfillment, Intent Fulfillment, User Satisfaction
Retrieval	Evaluating the retrieval step of a RAG pipeline independently	Retrieval Coverage, nDCG, Retrieval Precision
Other	General-purpose fallback for anything not covered above	General-purpose built-in properties
Custom	User-defined type built from scratch	You configure everything

Agentic Interaction Types

These types are used when evaluating multi-agent and agentic workflows. Each span in a trace is captured as a separate interaction, allowing Deepchecks to evaluate every component independently.

Interaction Type	Use case	Key properties
Root	The top-level span of a trace - the full end-to-end execution	Aggregated system metrics, overall quality
Agent	An AI agent that plans and delegates to tools or sub-agents	Plan Efficiency, Tool Coverage, Instruction Following, Tool Abuse
Chain	A multi-step chain that is not the root span	Aggregated child metrics
Tool	A tool called by an agent (web search, code execution, API call)	Tool Completeness
LLM	A direct LLM call within an agentic workflow	Reasoning Integrity, Instruction Following
Retrieval	A retrieval step within an agentic trace	Retrieval Coverage, Retrieval Precision

When using a supported framework (LangGraph, CrewAI, Google ADK, LangChain), each span is automatically captured and assigned the correct interaction type. You can also upload traces manually via the SDK.

See Uploading Agentic Data and Auto-Instrumentation for setup details.

Data fields

Each interaction type uses a different combination of content fields. The core fields (input, output) are needed for most property calculations. Additional fields unlock more evaluation capabilities:

Field	Required	What it contains
`input`	Required	The user's question, request, or source text
`output`	Required	The model's response or generated content
`full_prompt`	Recommended	The full prompt sent to the LLM, including system instructions. Required for Instruction Following and Instruction Fulfillment properties.
`information_retrieval`	Optional	Retrieved documents used as context (for RAG)
`history`	Optional	Conversation history or prior context
`expected_output`	Optional	Ground truth label or reference output
`steps`	Optional	Intermediate steps, tool calls, or span metadata (for agentic traces)

See Data Fields Reference for the full specification.