Grounding Score

Measures the exactness of claims against a gold-standard reference corpus. A high precision score indicates the model avoids introducing external knowledge or hallucinated details.

• Entailment Ratio: The percentage of generated claims logically entailed by the source.
• Contradiction Detection: Identifies statements that directly oppose the provided evidence.
• Granularity Mismatch: Penalizes outputs that are too vague or overly specific compared to the source.

Quantifies the proportion of verifiable claims in the output that are explicitly linked to a specific source passage. This is critical for auditability and right to explanation.

• Attribution Coverage: The ratio of cited sentences to total factual sentences.
• Source Fidelity: Verifies that the cited text actually supports the claim.
• Orphaned Claim Rate: The frequency of unsupported assertions appearing in the output.

Evaluates whether the model strictly answers the query using only the provided context, ignoring its parametric memory. This prevents data leakage and ensures the output is a function of the supplied documents.

• Instruction Following: Adherence to explicit directives like 'only use the provided text'.
• Out-of-Context Rejection: The model's ability to state 'I don't know' when the answer is absent.
• Distractor Resistance: Robustness against irrelevant passages injected into the context window.

Assesses whether the output captures all critical information from the source documents without omission. A grounded answer must be both faithful and exhaustive.

• Information Coverage: The percentage of key entities and relationships from the source present in the output.
• Semantic Overlap: Cosine similarity between the source embedding and the output embedding.
• Critical Fact Retention: Ensures safety-critical or legally binding clauses are never truncated.

Measures the stability of the grounding score when the source documents contain typos, OCR errors, or conflicting information. A reliable system maintains high fidelity despite adversarial data quality.

• Typo Tolerance: Score variance when character-level noise is injected.
• Conflict Resolution: The ability to synthesize a correct answer from contradictory sources.
• Format Invariance: Consistent performance across PDFs, HTML, and plain text.

Tracks the degradation of grounding over time as source documents become stale. This property is essential for continuous compliance monitoring in dynamic regulatory environments.

• Knowledge Cutoff Drift: The rate at which factual accuracy declines post-training.
• Versioned Grounding: The ability to anchor outputs to a specific document version or timestamp.
• Staleness Alerting: Automated triggers when a source document is updated, invalidating prior grounded outputs.

The foundational architecture that supplies the source documents against which a Grounding Score is measured. RAG injects retrieved chunks from a vector database into the model's prompt window, providing the factual context needed to answer a query. Without a robust retrieval step—typically using hybrid search combining semantic embeddings and keyword matching—the model has no external anchor, making grounding impossible. The quality of the retrieval pipeline directly sets the upper bound for the Grounding Score.

A standardized metric quantifying the frequency at which a model generates factually incorrect or nonsensical outputs. It is the inverse counterpart to the Grounding Score: a high grounding score implies a low hallucination rate. Benchmarks like TruthfulQA and HaluEval provide structured test suites to measure this. In enterprise settings, the hallucination rate is often calculated per-domain—legal, medical, financial—because a model grounded on general web text may still fabricate details when faced with proprietary jargon.

A specific evaluation measure that checks whether every claim in a generated output can be directly inferred from the provided source text. Unlike a general Grounding Score, which may incorporate relevance and completeness, faithfulness is a strict binary or graded assessment of entailment. A Natural Language Inference (NLI) model is typically used to detect contradictions. An output can be grounded but unfaithful if it introduces external knowledge not present in the source.

Two metrics that evaluate the quality of explicit source attribution in grounded outputs. Citation Precision measures what fraction of provided citations actually support the corresponding claim. Citation Recall measures what fraction of generated claims are backed by a citation. A high Grounding Score requires both: a model that cites irrelevant passages or makes unsupported assertions fails enterprise audit requirements. Tools like ALCE benchmark these capabilities.

The specialized storage system that indexes document embeddings for rapid semantic retrieval, forming the backbone of the grounding pipeline. Databases like Pinecone, Weaviate, and Milvus store high-dimensional vectors and enable approximate nearest neighbor (ANN) search. The choice of embedding model—such as text-embedding-3-large—and the chunking strategy directly impact the relevance of retrieved context and, consequently, the achievable Grounding Score.

A pre-generation filter that evaluates whether retrieved documents are actually pertinent to the user's query before they are passed to the language model. This prevents the model from being distracted by irrelevant chunks, a phenomenon known as context pollution. Low context relevance guarantees a low Grounding Score, as the model is forced to either ignore the noise or fabricate a bridge. This score is often calculated using a cross-encoder reranker.