TruLens — AI Glossary

Reference details

Related terms

• RAGAS
• LangSmith
• Langfuse

How it works

1. The application is instrumented so inputs, outputs and intermediate steps are recorded per call.
2. Quality is expressed as feedback functions — small scorers, heuristic or model-based, that return a value for a recorded step.
3. Because they are functions in your code, they are versioned and tested with the application rather than configured in a product.
4. For retrieval pipelines three scorers cover the common failure modes: whether retrieved context was relevant, whether the answer used it, and whether the answer addressed the question.
5. Those three isolate where a failure occurred, which is the thing a single end-to-end score cannot tell you.

Example: Three scores locate the failure that one cannot

An answer is judged poor. Context relevance low, groundedness high, answer relevance high means retrieval failed and generation faithfully used what it was given. Context relevance high with groundedness low means retrieval worked and the model ignored it. Context and groundedness both high with answer relevance low means it answered a different question. One aggregate score would read 'poor' in all three cases and point at nothing — the decomposition is the diagnostic.

Commonly confused with

An evaluation framework rather than a hosted platform, which is the distinction that matters against Langfuse, Braintrust or Phoenix: it gives you feedback functions to score RAG behaviour in code, not a service to send traces to. Its RAG triad — context relevance, groundedness, answer relevance — is the same decomposition RAGAS uses under different names.

When to use it

Reach for it when:

• Evaluation logic should live in your codebase, versioned with the application.
• A retrieval pipeline needs failures attributed to retrieval, grounding or answering.
• Custom scorers are needed beyond what a hosted product offers.

Reach for something else when:

• A managed dashboard and hosted tracing are what is actually wanted.
• Nobody will maintain the feedback functions, which rot like any other code.
• Prompt versioning and staged rollout are the primary need.

Continue your research

Use these connected ModelRefs sections to compare alternatives, inspect implementation paths, and review the evidence and governance boundaries relevant to TruLens — AI Glossary.

• Browse models
• Review providers
• Inspect benchmarks
• Explore workflows
• Learn AI concepts
• Read decision guides
• Review trust and methodology
• Start a decision path