How full document reading works
This guide explains what full document reading changes about how your agent reads your knowledge base, and when it applies.
When you ask your agent about your return policy, it normally searches your return policy document for the excerpts that answer your question, then pieces those together into an answer. Full document reading is a different way for your agent to use that same document: instead of searching it for excerpts, your agent reads the whole thing from start to finish, the way you would if you opened the file yourself.
How your agent normally answers a question
Your agent searches across your documents for the excerpts that best match your question. Those excerpts don't have to come from one place - a sentence from the top of a document, a paragraph from the bottom, even pieces from a different document entirely - and they're pulled based on how well they match your question, not in the order they appear in the original document.
This works well for most questions.
What's different when your agent already knows the document
Sometimes your agent already knows exactly which document has the answer. That's either because it found that document earlier in the conversation, or because you named it directly - for example, "What does the March report say about...". Your agent also weighs the rest of the conversation and your current prompt before deciding a full read is the right call, not just whether it can identify a document. When it decides to use full document reading, it reads that one document from the beginning, in order, instead of searching it for excerpts.
Two things follow from that:
- Nothing from other documents mixes in. Your agent is reading one document, start to finish, not searching your whole knowledge base, so nothing from unrelated documents can slip into the answer.
- The order stays intact. Since your agent reads the document the way it was written, things like numbered lists, step-by-step instructions, or specific wording stay exactly as they appear, instead of being pieced together out of order.
There's still a limit on how much it can read
Full document reading doesn't let your agent read an unlimited amount at once. Your agent's context depth setting still sets how much it can take in during one read, so if a document is longer than that setting allows, your agent starts from the beginning and stops once it reaches the limit. Refer to How context depth works to see how much that is for each plan.
When your agent still searches instead
Full document reading only happens when your agent already knows the one document to read. For a general question that could be answered from anywhere in your knowledge base, your agent still searches as normal. Full document reading works alongside search instead of replacing it.
Related articles
Updated about 1 hour ago