ADR-0005: Multi-Step Prompting for RAG Metrics

View Source

Status

Accepted

Date

2024-12-25

Context

RAG (Retrieval Augmented Generation) metrics like Faithfulness and Hallucination need to evaluate whether an LLM's output is grounded in the retrieved context. This requires:

  1. Understanding what claims the LLM made
  2. Understanding what facts exist in the context
  3. Comparing claims against facts
  4. Producing a score

A single-prompt approach ("Is this output faithful to the context? Score 0-1") produces inconsistent, poorly-calibrated scores with no explainability.

Decision

Implement multi-step prompting for RAG metrics, with each step producing structured intermediate outputs.

Example: Faithfulness metric

Step 1: Extract claims from actual_output
  → Returns: ["User logged in at 3pm", "Session lasted 2 hours"]

Step 2: Extract truths from retrieval_context
  → Returns: ["Login timestamp: 15:00", "Session duration: 120 minutes"]

Step 3: Generate verdicts comparing claims to truths
  → Returns: [
      {claim: "User logged in at 3pm", verdict: "yes", reason: "Matches login timestamp"},
      {claim: "Session lasted 2 hours", verdict: "yes", reason: "120 min = 2 hours"}
    ]

Step 4: Calculate score
  → Score = supported_claims / total_claims = 2/2 = 1.0

Consequences

Positive

  • Explainability: Each step produces inspectable intermediate results
  • Accuracy: Structured comparison more reliable than holistic scoring
  • Debugging: Clear visibility into why a score was assigned
  • Consistency: Deterministic scoring from verdicts (not LLM-generated scores)
  • Alignment with DeepEval: Matches the Python library's proven approach

Negative

  • Higher cost: 2-4 API calls per metric instead of 1
  • Increased latency: Sequential API calls add up
  • More prompts to maintain: Each step requires a carefully crafted prompt
  • Token overhead: Structured outputs include more tokens

Neutral

  • Intermediate results stored in Result.metadata for debugging
  • Each step uses structured outputs (JSON schema) for reliability
  • Prompt templates can be customized per-metric

Alternatives Considered

Single-prompt scoring

  • Rejected: Produces inconsistent scores without explanation. LLMs struggle to self-score accurately in a single pass.

Chain-of-thought in single prompt

  • Rejected: CoT reasoning is helpful but still produces unreliable numeric scores. Structured extraction is more robust.

Fine-tuned scoring models

  • Rejected: Requires training data and model hosting. Multi-step prompting works with any capable LLM.

Embedding similarity

  • Rejected: Semantic similarity doesn't capture logical entailment. "The sky is blue" and "The sky is not blue" have high similarity but opposite meanings.

References