Fact Verification

This initial processing stage breaks complex, multi-part statements into simpler, atomic claims that can be individually verified. Claim decomposition involves parsing a generated answer to isolate distinct factual assertions. Normalization then standardizes these claims into a canonical form (e.g., resolving coreferences like 'he' to named entities, standardizing dates) to enable precise matching against evidence. For example, the claim 'The CEO, who founded the company in 2012, later expanded to Asia' decomposes into: 1) 'The CEO founded the company.' 2) 'The founding year was 2012.' 3) 'The company expanded to Asia.'

This component searches a trusted knowledge base—such as a vector database, enterprise knowledge graph, or curated document store—to find source passages that support or refute each decomposed claim. It employs:

• Hybrid Search: Combines dense vector embeddings (for semantic similarity) with sparse lexical search (for keyword matching) to maximize recall.
• Multi-Hop Retrieval: For claims requiring synthesis, it performs sequential searches, using entities from initial results to query for related facts.
• Source Filtering: Prioritizes documents with high source reliability scores based on freshness, authoritativeness, and provenance.

The core reasoning module that determines the logical relationship between a claim and the retrieved evidence. A specialized NLI model (e.g., trained on datasets like FEVER or ANLI) classifies the claim as:

• Entailment: The evidence logically supports the claim.
• Contradiction: The evidence logically contradicts the claim.
• Neutral: The evidence provides insufficient information to verify the claim. This goes beyond simple keyword matching to assess semantic alignment and logical consistency, forming the basis for NLI-based verification.

This component creates a verifiable link between the system's output and its sources. It ensures answer grounding by:

• Generating Citations: Inserting inline references or footnotes that point to the exact evidence passages used for verification.
• Tracking Provenance: Maintaining an immutable audit trail that logs the original claim, all retrieved evidence documents, the NLI verdict for each, and the final composite verification score.
• Assessing Attribution Accuracy: Measuring whether the provided citations correctly and completely support the generated statement, a key metric for system evaluation.

This module synthesizes verification signals into actionable decisions. It calculates an overall confidence score for the original claim based on NLI results, evidence quality, and source reliability.

• Thresholding: Uses a predefined confidence threshold to categorize outputs as 'Verified', 'Unverified', or 'Contradicted'.
• Triggering Actions: Can initiate a refusal mechanism or generate an abstention signal for low-confidence claims, or trigger post-hoc verification or human review for borderline cases.
• Calibration: Aims to minimize calibration error, ensuring the confidence score accurately reflects the true probability of the claim being correct.

A closed-loop system that uses verification outcomes to improve its own components. It involves:

• Collecting Labels: Recording human corrections to system verdicts (e.g., false positives/negatives).
• Fine-Tuning: Using this feedback data to retrain the NLI verifier or evidence retriever, improving performance on domain-specific data (domain-adaptive retrieval).
• Metrics Monitoring: Continuously tracking key performance indicators like faithfulness metric scores, precision/recall of the retrieval engine, and attribution accuracy to detect regressions and guide improvements.

Faithfulness (also called factual consistency) measures the degree to which all factual claims in a generated answer are directly supported by and logically entailed by the provided source context. It is a critical metric for RAG systems.

• Evaluation Method: Typically assessed using a Natural Language Inference (NLI) model, where the generated claim is the hypothesis and the source context is the premise. The model classifies the relationship as entailment, contradiction, or neutral.
• Key Distinction: Different from answer relevance; a relevant answer can still be unfaithful if it introduces unsupported details.
• Example: If the context states "The company was founded in 2010," a faithful answer is "Founded in 2010." An unfaithful answer is "Founded in 2012."

Answer Relevance evaluates how directly the generated output addresses the original query, independent of its factual correctness. It ensures the system is not providing verbose, generic, or off-topic responses.

• Evaluation Method: Often measured by having an LLM or human judge score the answer on a scale (e.g., 1-5) based on its pertinence. Automated methods may embed the query and answer and measure their cosine similarity.
• Relationship to Faithfulness: An answer can be highly relevant but unfaithful (hallucinated details on-topic), or faithful but irrelevant (correct facts that don't answer the question).
• Example: For the query "What is the capital of France?", a relevant answer is "Paris." An irrelevant answer is "France is a country in Europe."

These retrieval-focused metrics assess the quality of the source material used for verification. They are prerequisites for high faithfulness.

• Contextual Precision: The proportion of retrieved document chunks that are actually relevant to answering the query. High precision means less noise in the context window.
• Contextual Recall: The proportion of all relevant document chunks in the knowledge base that were successfully retrieved. High recall ensures no critical evidence is missed.
• Impact on Verification: Poor recall leads to missing evidence, causing the model to guess (hallucinate). Poor precision burdens the model with irrelevant text, increasing the chance of distraction and erroneous synthesis.

Attribution Accuracy measures the correctness of the links (citations) between specific claims in a generated answer and the source passages that support them. It is the operationalization of source attribution.

• Evaluation Method: Requires granular, sentence-level or claim-level citations. Each citation is verified by a human or NLI model to confirm the source passage does indeed support the claim.
• Formula: (Number of Correctly Attributed Claims) / (Total Number of Attributable Claims).
• Importance: High attribution accuracy creates a verifiable audit trail, essential for debugging hallucinations and establishing trust in enterprise settings.

Self-Consistency is a sampling-based verification technique that leverages the stochastic nature of LLMs. The underlying assumption is that the most factually consistent answer will be generated most frequently across multiple attempts.

• Process: The same query is run multiple times (e.g., 5-10) with different random seeds, generating a set of candidate answers. The most frequent answer is selected as the final output.
• Use as a Metric: The agreement rate (e.g., 4 out of 5 runs produced the same answer) can serve as a proxy for confidence and factual stability. Low self-consistency signals high uncertainty or potential hallucination.
• Limitation: Computationally expensive and assumes the model's mode answer is the correct one.

Calibration Error quantifies the discrepancy between a model's predicted confidence score (e.g., a probability) for its answer and the actual empirical likelihood of that answer being correct. A well-calibrated model's confidence is a reliable indicator of truthfulness.

• Perfect Calibration: When a model says it is 80% confident, it should be correct 80% of the time.
• Common Metrics: Expected Calibration Error (ECE) and Maximum Calibration Error (MCE) are used to measure this. They bin predictions by confidence and compare the average confidence to the accuracy within each bin.
• Application to Verification: Enables reliable selective answering and refusal mechanisms. A model can be programmed to only answer when its calibrated confidence exceeds a defined confidence threshold (e.g., 90%).

The mechanism that links specific parts of a generated answer back to the exact document passages or data points used to produce them. This is the foundational capability for any verifiable RAG system.

• Key Implementation: Requires the model to generate citations (e.g., [1], [2]) inline with the text, corresponding to entries in a bibliography of retrieved chunks.
• Granularity Levels: Can range from document-level (citing a whole file) to sentence-level or phrase-level links for maximum auditability.
• Primary Benefit: Enables users and downstream systems to verify claims by consulting the original source material directly.

A quantitative measure of the degree to which a model's generated output is factually consistent with and supported by its provided source context. It is a critical benchmark for RAG evaluation.

• Common Techniques: Often calculated using Natural Language Inference (NLI) models to score if the answer is entailed by the context, or by comparing answer-context overlap via metrics like ROUGE or BERTScore.
• Industry Standard: Tools like RAGAS (Retrieval-Augmented Generation Assessment) incorporate faithfulness as a core metric alongside answer relevance and context precision.
• Purpose: Provides an objective score to compare different RAG architectures, prompting strategies, or fine-tuning approaches.

The use of Natural Language Inference models to determine if a generated claim is entailed by, contradicts, or is neutral to the provided source context. This provides a formal, model-driven check for factual alignment.

• Process: A separate, often smaller, NLI model (e.g., trained on datasets like MNLI or SNLI) takes the claim (from the LLM output) and the source context as a premise-hypothesis pair to classify their relationship.
• Architectural Role: Can be used as a post-hoc verification layer to filter or flag outputs, or integrated into the generation process to guide the LLM.
• Advantage: Offers a more nuanced check than simple keyword matching, understanding semantic equivalence and contradiction.

The process of adjusting a model's internal confidence scores (e.g., token probabilities) so they accurately reflect the true empirical probability of an output being correct. This enables reliable uncertainty quantification.

• The Problem: Modern LLMs are often poorly calibrated; a high softmax score does not reliably indicate a high chance of factual correctness.
• Techniques: Methods include temperature scaling, Platt scaling, or training with calibration-aware objectives.
• Use Case in RAG: A well-calibrated confidence score allows the system to implement selective answering, where it only responds when confidence exceeds a defined confidence threshold, otherwise triggering an abstention signal.

The process of breaking down a complex, multi-fact generated statement into simpler, atomic claims that can be individually verified against source material. This is essential for verifying sophisticated answers.

• Why It's Needed: A single paragraph from an LLM may contain several independent facts. Verification is more accurate when each atomic fact is checked separately.
• Method: Can be performed by prompting the LLM itself ("List the individual facts in this answer") or by using a dedicated decomposition model.
• Follow-on Step: Each atomic claim is then passed to a fact-checking module or NLI-based verification system for validation.

The systematic, immutable recording of the origin, lineage, and transformations of all data used to generate an output. This creates a complete forensic record for compliance, debugging, and trust.

• What It Includes: The original user query, the retrieved document IDs and chunks, the model prompt (with context), the raw generation, any verification scores, and the final formatted answer.
• Enterprise Critical: Essential for regulated industries (finance, healthcare) to demonstrate due diligence and for engineers to diagnose retrieval or generation failures.
• Implementation: Often built using structured logging, vector database metadata, and specialized observability platforms for AI pipelines.