Loading...
System Design and AI Engineering, by Roni Das. 980 interactive lessons.
Start LearningAI Engineering · Foundation · LLM and GenAI Ops
Requiring a model to answer from provided evidence and to cite it, rather than from its own memory. The core defense against hallucination in a RAG system.
In short
Grounding means making a model answer from evidence you provide, such as retrieved documents, and checking that each claim in the answer is supported by that evidence. It is the main defence against confident, unsupported answers in RAG systems.
In short
Grounding means making a model answer from evidence you provide, such as retrieved documents, and checking that each claim in the answer is supported by that evidence. It is the main defence against confident, unsupported answers in RAG systems.
Learn it in the course
This page is the free part.
₹999 in India$49 everywhere elseonce, for the whole course
The AI Engineering course covers Grounding across a run of lessons, not one page. These 4 alone are about 310 minutes of step-by-step reading, every one with code you run in the browser, all with a quiz.
Recall@k, context precision, and faithfulness split every wrong RAG answer into a retrieval bug or a generation bug, each with a different fix.
Why RAG fails in production as the corpus grows: a system that shines on a 100-document demo gets subtly worse at 5 million, offline metrics stay silent, and here are the levers that fight the RAG scale cliff.
A hands-on lab on two real models. An easy test set cannot tell a free model from a paid one, 30 near misses do more damage than 2,000 unrelated documents, and a missing prefix costs 20 points of recall.
RAG triples a chat bill through retrieved context. Cut it with a semantic cache, embedding reuse, and a reranker, without serving confidently wrong answers.
all part of
AI Engineering
210 lessons · about 219 hours · 10 free to read
₹999in India, by UPI
$49everywhere else, by PayPal
pay once,
yours for life
A language model can write a fluent answer from its training memory. That memory can be out of date, cannot be audited, and cannot be corrected by editing a document. Grounding means requiring the model to answer from evidence you supply, usually documents retrieved for the question, and being able to show which piece of evidence supports each claim.
A grounded answer is like a report with footnotes: every claim points to a source you can open. An ungrounded answer is like a confident colleague answering from memory. They may be right, but you cannot check, and when they are wrong they sound exactly the same.
Grounding has two parts. The first happens before the answer: put the right evidence in the prompt, and instruct the model to use only that evidence, to cite it, and to say when it does not contain the answer. The second happens after: check the answer's claims against the evidence, and block or regenerate the answer when they are not supported.
The standard measure is called faithfulness. The answer is split into single claims, because one sentence can hold two facts: "Refunds go to the original payment method and must be requested within 30 days" is two claims. Each claim is checked against the retrieved text by a checker, which can be a natural language inference (NLI) model, an LLM judge or a person. Faithfulness is the number of supported claims divided by the total.
Checking claim by claim is what makes the score useful. A bare score of 0.5 tells you to worry; a list of the two unsupported claims tells you what to fix. The RAGAS evaluation framework defines faithfulness this way and uses an LLM judge for the checks, so its scores carry the judge's own errors and are most trustworthy as comparisons between runs under the same judge.
Faithfulness is stricter than truth, on purpose. A claim that is true in the world but missing from the retrieved text counts as unsupported, because it came from the model's memory. Rewarding it teaches the system to bypass your documents, and on the day that memory is out of date, the same habit produces a confident wrong answer.
A grounding check compares the answer with the evidence the model was given. It cannot see evidence that never arrived. If retrieval missed the right document, a model can honestly summarise the wrong documents, score high on faithfulness, and still give a wrong answer. High faithfulness together with a wrong answer is the sign of a retrieval failure.
A measured run shows how wrong evidence pulls an answer away. Twenty questions were answered by the small model qwen3:4b and scored by how many of the right document's content words each answer contained. For a question about how big document chunks should be, the answer with no documents scored 12.7 percent. With the top 5 retrieved documents, which did not include the right one, it fell to 7.3 percent, because the answer built itself on a document about embedding storage. With the top 20, which did include it, it rose to 43.6 percent.
Where the evidence sits in the prompt matters too. With ten retrieved documents and the right one sixth, qwen3:4b identified it for 4 of 14 questions. With the same documents and the right one last, it identified it for 14 of 14. Grounding instructions cannot help if the model does not find the evidence it was given, so check recall and context order before you tune the prompt.
A grounding check is usually an output guardrail, a check that runs on the model's draft before the user sees it. It compares the draft's claims with the retrieved context and flags any claim the context does not support, so the answer can be regenerated or replaced. For a product built on retrieval, such as a documentation assistant, it is the most important output check.
It has a cost. Each check adds latency, and a stack that runs several checks one after another adds up quickly. Run cheap, independent checks in parallel, and keep the expensive ones for answers that need them. A grounding check can also be fooled by a plausible paraphrase, so treat it as one layer of defence, not the only one.
A grounded refusal is a feature. When the evidence does not contain the answer, "I could not find this, let me get a human" is better than a fluent guess. Build that path on purpose: tell the model to say when the documents do not answer the question, and count a grounded refusal as a correct result in your evaluation.
Diagnose in a fixed order. First recall: was the right chunk in the retrieved set? If not, fix retrieval, because no prompt can recover information that never reached the model. Then faithfulness: with the right chunk present, are there unsupported claims? If so, fix generation with grounding instructions, a claim-support gate or a different model. Then answer relevance: a fully grounded answer can still miss the question that was asked.
Keep your checker honest. If an LLM judge scores faithfulness, regularly compare a sample of its claim verdicts with human verdicts on the same answers, and track how often they agree. A judge upgrade or a small rubric edit can move that agreement without any warning.
Measure what the model knows without evidence. In a measured run, the same 20 questions answered with no documents at all reached about 44 percent of the score the pipeline with documents reached. If you only report the grounded score, you credit the documents for knowledge the model already had. Run your evaluation once with retrieval removed, and report the gap next to the level.
RAGAS
An open-source RAG evaluation framework that scores faithfulness by splitting an answer into claims and checking each against the retrieved context with an LLM judge.
ARES
A research evaluation system for RAG that trains lightweight judges and uses a small human-labelled sample to keep their scores calibrated.
Anthropic Claude Citations
A Claude API feature that returns citations pointing to the exact passages in supplied documents that support each part of an answer.
Google Vertex AI Grounding
A Vertex AI feature that grounds Gemini model answers in Google Search results or in your own data sources.
RAG is the pipeline that retrieves documents and puts them in the prompt. Grounding is the requirement that the answer comes from that evidence, plus the check that each claim is supported by it. A RAG system without a grounding check can still produce unsupported claims.
No. It catches claims the evidence does not support, but it cannot help when retrieval brought the wrong evidence, and a checker can be fooled by a plausible paraphrase. It is the strongest single defence, used together with good retrieval and human review for high-stakes answers.
Faithfulness is the share of an answer's claims that are supported by the retrieved context. The answer is split into single claims, each claim is checked against the context, and supported claims are divided by total claims.
Because faithfulness tests support by the retrieved documents, not truth in the world. A true claim missing from the documents came from the model's memory, which cannot be audited and goes out of date. Rewarding it teaches the system to ignore your documents.
The model is probably summarising the wrong documents faithfully. That is a retrieval failure: check whether the right chunk was in the retrieved set before you change the prompt.
RAG (Retrieval-Augmented Generation)
A pattern that retrieves relevant text from an external index and puts it in the model's prompt so the answer is grounded in current, verifiable facts rather than the model's memory.
Hallucination
When a language model produces fluent, confident text that is not supported by any source and is often simply false. Retrieval, grounding, and evaluation are the main defenses.
Guardrails
The checks around a model that screen its input for abuse and its output for grounding and safety, plus limits on what its tools can do. Containment, not just filtering, is what actually holds.
Agentic RAG
Agentic RAG is RAG where a language model runs the search in a loop. It decides whether to search, what to search for and where. Then it checks if the text it found answers the question.
Embedding
A list of numbers that represents the meaning of a piece of text, so that similar meanings sit close together in vector space. The same model must embed both the corpus and the query.
ANN (Approximate Nearest Neighbor)
Search that finds vectors close to a query without scanning every item, trading a small chance of missing a true neighbor for a large gain in speed. HNSW and IVF are the common indexes.
This page explains the idea. The "RAG Evaluation: Localizing Failure Between Retrieval and Generation" lesson builds it up step by step, with diagrams. It is part of the AI Engineering course, which has 210 lessons. 10 of them are free to read. You pay once and keep it for life, with no subscription.
Open the RAG Evaluation: Localizing Failure Between Retrieval and Generation lessoncourse 2
Take models from notebook to production, with labs on real models.