The Knowledge API: your own documents, cited, in one call

Most AI tools answer for you and leave you to trust it. The Knowledge API hands you the evidence first, and lets you decide who does the answering.

One call takes a question, searches the documents and conversations you already have in GPTree, and returns a block of passages with citation markers attached, sized to whatever context budget you name. You paste that into your own prompt, with your own model, in your own product.

Why it works this way

If you are building on top of an AI provider, you have probably already chosen your model. What you do not have is a trustworthy way to feed it your own material: retrieval you can inspect, passages you can verify, and a clean line between the evidence and the instructions.

That is what this returns. Every passage carries an identifier you can resolve back to its exact source and version. Every response tells you how the results were found and how fresh they are. When the index cannot serve a request properly, you get a clear failure instead of quietly worse results.

What a call gives you

• A ready-to-use context block with each passage marked, wrapped in clear boundaries so your model can tell evidence from instruction.
• A structured list of the passages used, with titles, sources, and relevance scores.
• Coverage numbers: how many passages were considered, how many made the budget, and how many were excluded because they no longer matched their source.
• Honesty stamps: how retrieval ran, how fresh the material is, and whether indexing is still catching up.

If you would rather it answered

There is a second endpoint that does write the answer, and it is deliberately opt-in: you have to authorize it with a separate kind of token, because generation costs money and reading does not.

What makes it worth using is what it refuses to do. Every citation marker in the answer points at a real passage you can open; markers the model invents are stripped out before you ever see them and counted so you know it happened. If the model answers without citing anything it was given, you get an honest refusal rather than a confident paragraph with nothing behind it. And when nothing in your knowledge matches the question, no model is called at all and nothing is charged.

Every answer is also recorded: which passages it used, how fresh they were, and which model wrote it. That record is what will later let GPTree tell you when an answer you relied on has gone stale because its source changed.

What it still does not do

It does not reach into your chat conversations to answer on your behalf, and it never blends the two questions that matter most: how well supported an answer is, and how much of your knowledge was searchable in the first place. Those stay separate fields, because collapsing them is how confident wrong answers get made.

Getting started

Create a read-only token in Account, then call the search or context endpoint with your question. A token carries only the permissions you granted it: reading, asking, or adding documents. A read-only token cannot send messages, spend credits, or reach anything else in your account. The full endpoint reference, including request fields and status codes, is at the Knowledge API documentation.

Reading and asking are available on every plan, including Free, and so is sending your own documents in as text you have already extracted. Uploading a file for GPTree to read, and connecting an assistant over MCP, are on Pro and Team. Rate limited per account, and scoped to material you own.