To reduce hallucinations in LLMs, you can craft prompts that explicitly constrain the model to use only the information you provide. Start by setting clear boundaries in the instruction, such as "Answer strictly based on the provided data, and do not add external knowledge." For example, you might structure the prompt with a preamble like: "You are an assistant that answers questions using only the information provided below. If the answer isn't in the text, say 'I don't have enough information.' Do not guess or infer details." This creates a guardrail by forcing the model to prioritize the provided context and admit uncertainty. Including specific examples of desired behavior (e.g., showing how to handle missing data) reinforces the rule.
Next, structure the input to separate the context from the query clearly. Use markers like "Context:" and "Question:" to avoid ambiguity. For instance: "Context: [paste your data here]. Question: [ask your question here]. Answer using only the context. If unsure, say 'Not specified in the context.'" This format helps the model recognize the scope of allowed information. You can also add warnings like "Do not invent names, dates, or facts" to block common hallucination patterns. For technical topics, include domain-specific constraints (e.g., "Use only the API documentation provided; do not reference other frameworks").
Finally, test and iterate. If the model still hallucinates, add escalating penalties in the prompt, such as: "If you include information not in the context, your answer will be rejected." Combining these techniques reduces overcreativity while maintaining utility. However, no method is foolproof—always validate outputs against source material, especially for critical use cases like medical or legal advice.