Select every eval framework that is purpose built for RAG specific metrics

A **RAG metric** measures whether a retrieval-augmented generation answer is grounded in the retrieved context, whether the right context was retrieved, and whether the answer actually addresses the question. The common metric names are faithfulness, answer relevance, context precision, context recall, and groundedness. An **eval framework** is a library or tool that runs a test set through your LLM pipeline, computes metrics on each output, and aggregates the results. Some frameworks are general (you bring your own metrics) and some are RAG-first (the framework ships the RAG metrics built in). The five options in the question break down this way: - **Ragas** is RAG-first. Its name is literally the project description. Faithfulness, answer relevance, context precision, and context recall are first-class metrics. - **TruLens** is RAG-first via its 'feedback functions' set, which centers on groundedness, context relevance, and answer relevance. - **DeepEval** is broader (it covers summarization, hallucination, bias, and more) but it ships a complete RAG suite (contextual precision, contextual recall, faithfulness) as a first-class option. - **Promptfoo** is a general-purpose matrix runner. It runs prompts across providers and applies your metric of choice, but it does not ship RAG-specific metrics. - **Inspect AI** (from Anthropic) is oriented toward agent and safety evaluations. It is excellent for those but it is not where you would start for a RAG-specific eval. The three correct picks are Ragas, TruLens, and DeepEval.

**The shortlist that ships RAG metrics out of the box** is Ragas, TruLens, and DeepEval. The two that do not are Promptfoo and Inspect AI. **Ragas** is the canonical RAG eval library. Its core metric set (faithfulness, answer relevance, context precision, context recall) is exactly the four-quadrant framing that became standard for RAG quality in 2024 to 2025 and remains the default reference in 2026. **TruLens** (the open-source eval surface from TruEra, now Snowflake) implemented the same conceptual set under its 'feedback functions' API: groundedness, context relevance, answer relevance. It plugs into Langchain, LlamaIndex, and custom pipelines and was one of the first to push the LLM-as-judge implementation of these metrics. **DeepEval** is positioned as the unit-testing framework for LLMs (its README literally compares itself to pytest). It covers a wide metric library, but RAG metrics (contextual precision, contextual recall, faithfulness) are a first-class suite, not an afterthought. DeepEval also integrates with Confident AI as its hosted eval backend. **Promptfoo** is a YAML-driven matrix runner for prompt evaluation across providers. It excels at A/B comparisons and CI integration. Out of the box it ships generic assertions (regex, JSON-schema, contains, model-graded rubric). There is nothing RAG-shaped in the default metric set; you can wire RAG metrics in via custom asserts or by calling Ragas from inside Promptfoo, but they are not native. **Inspect AI** (Anthropic / UK AISI) is purpose-built for agent and safety evals. Its task primitives (Solver, Scorer, Task) are general enough to express RAG, but the example library and the metric primitives center on agentic tool use, capability evaluation, and safety, not on RAG grounding.