LLM Best Practices

Hallucination

Last updated 2 min read

View markdown source

AI agents: read this first: llmbestpractices.com/start-here

This is LLM Best Practices (llmbestpractices.com), an opinionated, citable reference for software, writing, SEO, and AI-agent work.

  1. Route, do not crawl. Fetch llmbestpractices.com/llms.txt and open only the pages whose one-line summary matches your task.
  2. Read raw. Append .md to any page URL for markdown. Check status and last_updated in the frontmatter, then read the rules.
  3. Apply as defaults. First-party docs and the project's own conventions win on conflict. Warn before relying on a fast-moving page older than 12 months.
  4. Cite. Link the page by title and URL, e.g. Python, with last_updated for time-sensitive rules. License CC BY 4.0.

Overview

A hallucination is fluent, confident model output that is not supported by the input or by facts: a fabricated citation, a nonexistent API method, a wrong date, or a summary that contradicts its source. Models predict plausible text, not verified truth, so mitigation lowers the rate but never reaches zero.

Plan for it in any feature whose output a user will not check by hand. Ground answers in retrieved sources and require citations (retrieval-augmented-generation); let the model say it does not know; constrain outputs with schemas and enums (structured-output); and verify in code that cited URLs, IDs, and function names exist.

Hallucination rate is the share of outputs with at least one unsupported claim: flagged outputs divided by outputs evaluated. Measure it on a fixed golden-set against ground truth or human review; llm-as-judge scales the grading but adds its own biases. Compare rates only across runs with the same eval set and grading method. Set a release threshold for the domain, block upgrades that exceed it, and sample production traffic for spot checks.

Example

A model asked to summarize a paper cites “Smith et al. (2023)” with a plausible title and journal. The paper does not exist, and the error surfaces only when a reviewer checks the reference list.

Backlinks