LLM learning

LLM Limitations & Grounding

Recognize unsupported output and design evidence, citations, abstention, and evaluation.

Not started3 min explanation

Visualize, practice, and deep-dive material are optional—use only what helps you learn.

Explanation

A focused 3-minute explanation using the topic's authored material.

Learning goals and prerequisites

After this lesson

  • Classify important LLM failure modes
  • Map accepted claims to evidence
  • Evaluate answerable, conflicting, and unanswerable cases

Helpful before starting

  • LLM prompting
  • Basic source and citation literacy

Mental model

Grounding connects every claim to support

Direct answer

Grounding gives a model identifiable evidence or verified tool results and accepts an answer only when its claims are supported by that evidence. When the required support is missing or conflicting, the system should clarify, answer partially, abstain, or escalate instead of filling the gap with plausible text.

Follow the mechanism

  1. Retrieve sources the user is authorized to access and preserve their identity, version, and freshness.
  2. Build a bounded context without separating passages from their provenance.
  3. Generate an answer whose claims cite the relevant source spans.
  4. Verify claim-to-evidence support and apply the configured abstention or escalation policy before release.

Running example

The available passage says: “Refunds are accepted within 30 days with a receipt.” For “How long do I have to request a refund?”, the answer “You can request a refund within 30 days if you have a receipt” is supported and can cite refund-policy. For “Is international shipping free?”, the same source provides no support. A fluent “yes” may sound helpful, but the correct grounded result is “I do not have evidence to answer that question.” The missing answer is a successful safety behavior, not an incomplete generation.

What this does not mean

Displaying a citation handle does not prove that the source supports the nearby claim. Retrieval can find the wrong passage, context construction can omit the right one, generation can contradict good evidence, and a validator can accept a citation that merely exists. Evaluate those stages separately so the correction targets the actual failure.

Key points

  • Authorized retrieval → bounded context → cited claims → support verification → release policy is the grounding chain.
  • A citation is useful only when the cited span supports the associated claim.
  • Answerable, unanswerable, stale, and conflicting cases require separate evaluation fixtures.
  • Abstention is correct output when the evidence contract cannot be satisfied.
Try it: Predict whether the evidence can answer

Use the refund passage for both questions before advancing the grounded-answer pipeline.

  1. 1.Underline the exact span that supports the refund answer.
  2. 2.Predict what the system should release for the shipping question and explain why.
  3. 3.Advance the pipeline and identify whether each visible failure belongs to retrieval, context construction, generation, citation, or policy.

Was this lesson helpful?

Submit to the team when server feedback is available; otherwise this browser keeps a local copy and tells you so.

Back to Topic Library