# Hallucinations

> Language model answers that sound convincing but aren't true: made-up numbers, links, quotes or API methods

Источник: https://pachca.com/en/glossary/hallucinations

Hallucinations are errors where a language model confidently gives you wrong information. A model generates the most likely text rather than checking facts, so it can invent a source, an API method that doesn't exist or a number that isn't in your docs.

Why hallucinations happen:

- The data wasn't in the training set or the context
- The question is ambiguous
- The model fills in the answer based on similar examples
- The context is overloaded and key details get lost

How to reduce them:

- Give the model sources through RAG and ask it to cite them
- Connect accurate docs instead of relying on the model's memory
- Have a person review important answers
- Let the model say “I don't know”

When an agent works with the Pachca API, it doesn't need to guess methods: up-to-date docs are available in llms.txt, in Agent Skills and in the OpenAPI spec, and Pachca's MCP server gives the agent an exact description of every tool.

For developers

[Docs on dev.pachca.com →](https://dev.pachca.com/guides/ai-agents/overview?utm_source=pachca.com&utm_medium=glossary&utm_campaign=dev-portal)