Hallucination Filter

The core mechanism that cross-references generated statements against a vector database or knowledge graph in real-time. Before outputting a claim, the filter queries a trusted corpus, retrieves semantically relevant documents, and computes an entailment score to determine if the generated text is supported by the evidence.

• Uses dense retrieval with embedding similarity
• Computes Natural Language Inference (NLI) scores
• Flags contradictions and unsupported assertions

Assigns a numerical confidence score to each generated sentence based on its alignment with the retrieval corpus. Sentences falling below a calibrated entailment threshold are either rewritten or suppressed entirely.

• Typical thresholds: 0.7–0.9 entailment probability
• Uses models like DeBERTa or T5 fine-tuned on NLI tasks
• Enables granular, sentence-level intervention without blocking the entire response

Identifies when a generated statement directly conflicts with retrieved evidence or when two generated sentences contradict each other. The filter applies logical consistency checks across the full output span.

• Detects intra-output contradictions (e.g., 'The patient has a fever' followed by 'The patient is afebrile')
• Detects output-corpus contradictions against source documents
• Triggers regeneration or suppression of conflicting spans

Measures the model's epistemic uncertainty about generated claims. When the model lacks sufficient grounding evidence, the filter forces an explicit abstention or inserts a calibrated uncertainty disclaimer rather than outputting speculative content.

• Uses predictive entropy and mutual information metrics
• Applies conformal prediction to guarantee coverage at a specified confidence level
• Prevents confident-sounding fabrications in low-evidence scenarios

Enforces that every factual claim includes a verifiable citation to a source document in the retrieval corpus. The filter strips unsupported assertions and appends provenance links, enabling human auditors to trace each statement back to its origin.

• Maps generated spans to specific document chunks
• Inserts inline citations or footnote references
• Critical for regulated industries requiring audit trails

Implements a defense-in-depth strategy by chaining multiple heterogeneous filters. A lightweight regex and keyword scanner catches obvious fabrications, a neural entailment model scores semantic consistency, and a final LLM-as-a-Judge resolves ambiguous cases.

• Reduces false negatives through overlapping coverage
• Optimizes latency by routing simple cases to fast classifiers
• Escalates edge cases to more computationally intensive review

The foundational architecture that supplies the grounding corpus for a hallucination filter. RAG retrieves relevant documents from a vector database in real-time, providing the factual evidence against which generated statements are verified. Without a high-quality retrieval step, the filter lacks a source of truth.

• Semantic search identifies contextually relevant chunks.
• Hybrid retrieval combines keyword (BM25) and vector search.
• The filter compares generated tokens to the retrieved context to flag unsupported spans.

A prompting architecture where the LLM acts as its own internal hallucination filter. The model generates an initial response, then evaluates its own output for factual consistency against the provided context before presenting it to the user.

• Uses chain-of-verification to generate targeted questions.
• Identifies self-contradictions within long-form outputs.
• Often combined with external filters for defense-in-depth.

A runtime inference technique that enforces strict schema adherence by masking invalid tokens. While primarily used for structured output, it acts as a syntactic hallucination filter by preventing the generation of malformed JSON, invalid API calls, or out-of-vocabulary terms.

• Applies a logit bias to suppress disallowed tokens.
• Forces outputs to conform to a predefined grammar or regex.
• Eliminates structural hallucinations in code generation tasks.

A defense-in-depth architecture that combines multiple heterogeneous safety classifiers. A hallucination filter often operates as one node in this ensemble, working alongside toxicity detectors and jailbreak classifiers to provide comprehensive output validation.

• Uses a voting or cascading mechanism to reduce false negatives.
• Combines regex, semantic embedding, and neural classifiers.
• The hallucination filter specifically handles factual grounding, while other guards handle policy violations.

An alternative grounding source to unstructured text corpora. Instead of verifying against retrieved documents, the hallucination filter validates generated entity relationships against a deterministic knowledge graph. This provides high-precision factual verification for structured domains.

• Checks if a predicted triple (subject, predicate, object) exists in the graph.
• Ideal for biomedical, legal, and financial ontologies.
• Eliminates contradictions with established organizational facts.

An evaluation paradigm where a strong, general-purpose model scores the factual accuracy of another model's output. This serves as a scalable, automated hallucination filter for evaluation pipelines, replacing human review for high-volume monitoring.

• Provides a faithfulness score comparing output to source context.
• Used in offline evaluation rather than real-time blocking.
• Enables continuous monitoring of hallucination rates in production.