Give it the source.
Ask from a source, not from the model's memory. Point it at the page, ask for the sentence the answer rests on, and let it say the page does not say.
new words
- Source
- The file or page an answer comes from, which you can open and read yourself.
- Knowledge cutoff
- The date after which the model read nothing. Your company's news is after it.
- Hallucination
- A fluent answer that is false: the model filled a gap with what sounds right.
the lesson
Ask a model a question with nothing to read, and it answers from its training. That text is old, it is everyone's, and it names no source. When it has no answer, it can still write one that sounds right.
Three habits close that gap:
| Habit | Why |
|---|---|
| Name the file or page | The model reads your text, not its memory of the world. |
| Ask for the sentence | A quoted line can be checked in a second. A paraphrase cannot. |
| Allow "not in the file" | A model told it may say so is less pushed to fill the gap. |
Retrieval systems build the same rule into their prompts: passages first, the question last, and "answer only from these, or say so": RAG, lesson 6.
With instibrain, your AI reads your pages before it works and names the page an answer comes from. Without it, name the file in your message, or paste the passage above your question.
try it
Pick a file in the folder you started in. Type this, with its name, and press Enter:
Read <file>. What is it for? Quote the sentence that says so. If the file does not say, say so.
you should see
- An answer with a sentence quoted from the file, which you can find in it.
- Or a plain "the file does not say", which is a right answer too.