[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"glossary-grounding::en":3,"gloss-cluster-grounding::en":20,"gloss-next-grounding::en":9},{"slug":4,"category":5,"name":6,"definition":7,"meta_desc":8,"faq":9,"schema_markup":9,"related":10},"grounding","core-ai","Grounding","Grounding is the practice of anchoring an LLM's output to specific, verifiable, current source data rather than letting it rely purely on the general (and potentially outdated or imprecise) knowledge baked into its training weights. A model's parametric knowledge — everything it \"knows\" from training — is inherently fuzzy, dated as of its training cutoff, and prone to conflation and hallucination on specifics; grounding fixes this by explicitly supplying the exact facts the model needs at inference time and instructing it to base its answer on that supplied material rather than what it might otherwise recall. This matters enormously for SaaS builders because it is the practical difference between an AI feature that's an impressive demo and one that's trustworthy enough to ship to customers who will act on its output — pricing pages, account-specific data, current inventory, today's exchange rates, or your company's actual refund policy are all things an ungrounded model will either not know accurately or will guess at with false confidence. The primary grounding technique is retrieval-augmented generation (RAG): retrieving relevant, current documents and injecting them into the prompt so the model answers from that specific context rather than memory. But grounding also includes simpler techniques: directly injecting known facts into the system prompt (e.g., \"Today's date is 2026-07-02. Our current refund window is 30 days.\"), providing function-calling access to live data sources (letting the model query a database or API for the current answer rather than guessing), and instructing the model explicitly to say \"I don't know\" or defer when the supplied context doesn't contain the answer, rather than filling the gap with a plausible-sounding guess. A concrete worked example: an internal HR chatbot answering \"how many vacation days do I have left?\" must never answer from the model's general training knowledge (which has no idea about this specific company's specific employee's specific balance) — it must be grounded via a function call to the actual HR database, returning the real number, with the model's job reduced to phrasing that real number into a natural sentence rather than generating the number itself. Ungrounded AI features that \"sound right\" but aren't tied to real data are the most common source of embarrassing, trust-destroying AI failures in production SaaS. A useful mental test for whether a given AI feature needs grounding: ask \"could this answer plausibly be wrong in a way the user wouldn't notice, and would that matter?\" Features answering general knowledge questions with low stakes (\"explain what REST APIs are\") can lean more on the model's parametric knowledge; features answering specific, actionable, or account-relevant questions (\"what's my current invoice total,\" \"is this contract clause standard\") need explicit grounding via retrieval or function calls, because an ungrounded wrong answer in those contexts causes real user harm or business risk, not just embarrassment.","Grounding constrains an LLM's output to verifiable source data, rather than its general training knowledge, to reduce hallucination.",null,[11,14,17],{"slug":12,"name":13},"function-calling","Function Calling (Tool Use)",{"slug":15,"name":16},"hallucination","Hallucination",{"slug":18,"name":19},"retrieval-augmented-generation","Retrieval-Augmented Generation (RAG)",[21,25,29,33,36,39,42,45,48,51,54,57],{"slug":22,"category":5,"name":23,"updated_at":24},"agentic","Agentic AI","2026-08-24T02:46:36+00:00",{"slug":26,"category":5,"name":27,"updated_at":28},"alignment-tax","Alignment Tax","2026-08-24T02:46:37+00:00",{"slug":30,"category":5,"name":31,"updated_at":32},"artificial-intelligence","Artificial Intelligence (AI)","2026-08-24T02:46:38+00:00",{"slug":34,"category":5,"name":35,"updated_at":24},"attention","Attention",{"slug":37,"category":5,"name":38,"updated_at":32},"beam-search","Beam Search",{"slug":40,"category":5,"name":41,"updated_at":28},"benchmark-contamination","Benchmark Contamination",{"slug":43,"category":5,"name":44,"updated_at":28},"catastrophic-forgetting","Catastrophic Forgetting",{"slug":46,"category":5,"name":47,"updated_at":32},"computer-vision","Computer Vision",{"slug":49,"category":5,"name":50,"updated_at":28},"constitutional-ai","Constitutional AI",{"slug":52,"category":5,"name":53,"updated_at":24},"context-window","Context Window",{"slug":55,"category":5,"name":56,"updated_at":32},"deep-learning","Deep Learning",{"slug":58,"category":5,"name":59,"updated_at":24},"diffusion-model","Diffusion Model"]