Lecture on LLM Hallucinations

Introduction

• Focus on LLM hallucinations, detection, and mitigation.
• Emphasis on building Retrieval-Augmented Generation (RAG) systems.
• Discuss metrics from the RAGAS paper for evaluating RAG pipelines.

Understanding Hallucinations

• Hallucinations in LLMs: False/misleading information due to statistical nature.
• Example: Air Canada's chatbot gave false refund policy information.
• No complete solution for hallucinations; they occur because LLMs follow probability, not truth.
• Mitigation possible when facts are available for grounding LLM responses.

RAG Pipelines

• Retrieve document chunks relevant to a user's query.
• LLM uses retrieved context to generate a response, reducing hallucination chances.
• Checking faithfulness or groundedness of LLM's output against retrieved context.

Metrics for RAG Systems (RAGAS Framework)

• Faithfulness: Is the LLM's output grounded in the provided context?
• Context Relevance: Effectiveness of retrieval—are retrieved documents relevant to the query?
• Answer Relevance: Does the generated answer address the user's question directly?

Methods for Evaluating Metrics

• Sentence Embedding Models: Measure semantic similarity between text snippets.
• LLMs as Judges: Prompt LLMs to score and evaluate the faithfulness of responses.
  • Example: Lynx model for hallucination detection in specialized domains.

Detailed Evaluation Techniques

• Faithfulness Checks:
  • Use sentence embeddings and cosine distance to assess similarity.
  • LLMs as judges to score answer's faithfulness.
  • Two-step method: Break down answer into factual statements and verify each against context.
• Answer Relevance:
  • Generate questions from the answer and compare them with the original question using embeddings.
• Context Relevance:
  • Extract key sentences relevant to the question from retrieved context.
  • Measure ratio of relevant to total sentences.

Reference-Free Metrics

• Metrics do not require a correct reference answer, suitable for real-time application.

Application of Metrics

• Real-Time Use:
  • Check faithfulness before displaying answers to users.
  • Show message if answer can't be generated faithfully.
• Development Phase:
  • Test RAG pipeline with user questions.
  • Evaluate and improve with faithfulness, context, and answer relevance metrics.

Conclusion

• Hallucinations are a product of LLMs' statistical nature but can be managed.
• Faithfulness checks and RAGAS framework metrics are essential for improving RAG systems.
• Invitation to subscribe for further insights and controls in practice.