LLM Hallucination Mitigation: Practical Techniques for Production Systems

The Hallucination Problem Is Not Going Away

Despite massive improvements in LLM capabilities, hallucination remains the single biggest barrier to enterprise AI adoption. Models confidently generate plausible-sounding but factually incorrect information. In production systems where accuracy matters -- healthcare, legal, financial services -- even a 2% hallucination rate can be unacceptable.

The reality is that hallucination is an inherent property of how LLMs work. They generate text based on statistical patterns, not by reasoning over verified facts. Mitigation, not elimination, is the practical goal.

Technique 1: Retrieval Grounding (RAG)

The most widely adopted mitigation strategy. Instead of relying on the model's parametric knowledge, retrieve relevant documents and include them in the context:

# Simplified RAG pipeline
documents = vector_store.similarity_search(user_query, k=5)
context = "\n".join([doc.content for doc in documents])

response = llm.generate(
    system="Answer based ONLY on the provided context. "
           "If the context doesn't contain the answer, say so.",
    messages=[{
        "role": "user",
        "content": f"Context: {context}\n\nQuestion: {user_query}"
    }]
)

RAG reduces hallucination by giving the model a source of truth, but it does not eliminate it. Models can still hallucinate details not in the retrieved documents or misinterpret the retrieved content.

Technique 2: Structured Output with Schema Validation

Constraining the model's output to a strict schema prevents entire categories of hallucination:

from pydantic import BaseModel, Field
from enum import Enum

class Confidence(str, Enum):
    HIGH = "high"
    MEDIUM = "medium"
    LOW = "low"

class FactualClaim(BaseModel):
    claim: str
    source_document: str = Field(description="Which retrieved document supports this claim")
    confidence: Confidence
    direct_quote: str = Field(description="Exact quote from source supporting the claim")

By requiring the model to cite specific sources and provide direct quotes, you create an auditable chain from claim to evidence.

Technique 3: Chain-of-Verification (CoVe)

A multi-step approach where the model verifies its own output:

1. Generate: Produce an initial response
2. Plan verification: Generate a list of factual claims that need checking
3. Execute verification: For each claim, independently verify it against the source material
4. Revise: Produce a final response that removes or corrects unverified claims

Research shows CoVe reduces hallucination rates by 30-50% compared to single-pass generation.

Technique 4: Confidence Calibration

LLMs are notoriously poorly calibrated -- they express high confidence even when wrong. Techniques to improve calibration:

• Verbalized confidence: Ask the model to rate its confidence (1-10) for each factual claim and filter low-confidence claims for human review
• Consistency sampling: Generate multiple responses at non-zero temperature and flag claims that appear in fewer than 80% of samples
• Logprob analysis: Examine token-level log probabilities to identify when the model is uncertain (available with some APIs)

Technique 5: Guardrail Layers

Deploy post-generation validation:

• NLI-based fact checking: Use a Natural Language Inference model to check whether generated claims are entailed by the source documents
• Entity verification: Extract named entities from the response and verify they exist in the source material
• Numerical validation: Check that any numbers, dates, or statistics in the response match the source data

Production Architecture Pattern

The most reliable production systems layer multiple techniques:

1. Retrieve relevant documents (RAG)
2. Generate response with structured output schema requiring source citations
3. Run NLI-based entailment check against retrieved documents
4. Flag low-confidence or unverified claims
5. Route flagged items to human review queue

This layered approach typically achieves 95%+ factual accuracy in domain-specific applications, compared to 70-80% with naive prompting.

Metrics to Track

• Groundedness score: Percentage of claims supported by retrieved documents
• Faithfulness: Whether the response accurately represents the source material (not just supported by it)
• Hallucination rate: Percentage of responses containing at least one unsupported claim
• Abstention rate: How often the system correctly says "I don't know" instead of hallucinating

Sources: Chain-of-Verification Paper | RAGAS Evaluation Framework | Vectara Hallucination Leaderboard