grounding

Grounding ensures AI outputs are anchored in retrieved facts rather than generated from potentially unreliable model knowledge. **Problem**: LLMs may generate plausible but false information from training data or hallucination. Grounding constrains outputs to verified sources. **Mechanisms**: **Explicit grounding**: Only answer from retrieved context, refuse if information not found. **Soft grounding**: Prefer retrieved info, mark uncertain claims. **Verification**: Check outputs against sources, flag unsupported statements. **Implementation**: System prompts emphasizing only using provided context, retrieval-augmented generation, post-generation verification against sources. **Grounding indicators**: Confidence scores, source citations, explicit uncertainty markers ("According to...", "The document states..."). **Trade-offs**: May refuse valid questions if retrieval fails, reduced creativity/synthesis. **Enterprise use**: Critical for compliance, legal liability, accurate customer support. **Google's approach**: Grounding API connects Gemini to Google Search for real-time factual grounding. **Best practices**: Clear grounding policies, handle "information not found" gracefully, combine with retrieval quality optimization. Foundation of trustworthy AI assistants.