How to Make AI Smarter by Secretly Adding Facts to Your Questions

This patent describes a system that automatically finds important words in your questions to an AI, then adds relevant facts from a knowledge base before the AI answers, making the AI's response more informed without you knowing the extra facts were added.

What it covers

This system improves how large language models (LLMs) respond by automatically adding extra information to a user's question. First, it receives user input, like a question or statement (Claim 1). It then analyzes this input to find key 'entities' — specific people, places, or things mentioned (Claim 1). Using a 'semantic framework,' which is like a structured database of facts (Claim 4, e.g., a knowledge graph), it looks up information related to these entities. This information is often in the form of 'triples' (subject, predicate, object, like 'Apple Inc. designs iPhones') (Claim 1). These triples can be turned into plain sentences, called 'verbalized triples' (Claim 3). The system then adds these extra facts to the original user input, creating a 'modified user input' (Claim 1). This modified input is sent to the LLM, which uses the added facts to generate a more complete response. Crucially, the user only sees the original question and the LLM's final answer; the added facts are hidden (Claim 9). For example, if a user asks, 'Tell me about the CEO of Apple,' the system might secretly add facts like 'Tim Cook is the CEO of Apple Inc.' and 'Apple Inc. is a technology company' to the prompt before the LLM processes it.

What it doesn't cover

• —Does not cover systems where the user explicitly provides all the supplemental information to the large language model.
• —Does not cover systems that only use a large language model without an external semantic framework or knowledge graph to find additional facts.
• —Does not cover methods where the modified user input, including the added facts, is displayed to the user before being sent to the LLM.
• —Does not cover systems that only identify entities in the user input without then using those entities to retrieve and add 'triples' or similar structured data.
• —Does not cover general prompt engineering techniques that manually add context or instructions to an LLM prompt.

The clever bit

The novelty lies in the automated, hidden enrichment of user prompts with structured, factual data from a semantic framework, ensuring the LLM receives comprehensive context without user intervention or awareness of the modification.

Why it matters

This patent addresses a core challenge with large language models: their tendency to 'hallucinate' or provide incorrect information, especially on specific, up-to-date facts. By automatically injecting verified data from a knowledge graph, it aims to make LLM responses more accurate and reliable. This approach is vital for enterprise applications where factual correctness is paramount, such as in business intelligence, customer support, or legal research tools. It allows LLMs to leverage curated, external data without requiring users to manually provide extensive context.

Real-world examples

• 1.Enterprise AI assistants using internal company knowledge bases
• 2.Customer support chatbots providing detailed product information
• 3.Research tools that augment user queries with academic or scientific data
• 4.Legal AI systems that inject case law or statute details into user questions
• 5.SAP's own AI offerings that integrate with business data