Knowledge
Making a flow answer from your own documents, and the order to build it in so you can tell where it went wrong.
A model on its own answers from what it learned. To answer from your material, a flow needs a retrieval step: something that finds the relevant passages first and hands them to the model with the question.
Two places hold material in Studio, both at the bottom of the left rail on the home screen:
- Knowledge — document sets a flow can search.
- My Files — files you have uploaded, whether or not a flow uses them.
The shape of a retrieval flow
The Vector Store RAG template is this pipeline already assembled. If you build it yourself, build it in this order — and run it after each step:
- Load the documents.
- Split them into chunks. Too large blurs the answer; too small loses the context that made the passage meaningful.
- Embed and store, so passages can be searched by meaning rather than by keyword.
- Retrieve the closest chunks for the question.
- Prompt the model with the question and those chunks.
Run it after step 4, before adding the model. Look at what retrieval actually returned. If the wrong passages come back, no prompt will rescue the answer — and a retrieval problem is far easier to see on its own than through a model's paraphrase of it.
Chunking, in practice
The first thing to change when answers are vague is chunk size, not the prompt. A chunk should be big enough to carry a complete thought and small enough that most of it is relevant to any question it matches.
For prose, paragraphs are a reasonable unit. For reference material with headings, splitting on headings usually beats splitting on length.
Cost and time
Embedding runs once per chunk, so the first ingest of a large document set is slow and everything after it is fast. That first run is also where the cost sits. Ingest a representative sample before committing a whole corpus — it is much cheaper to discover a chunking mistake on ten documents than on ten thousand.
What this is not
This is Studio's retrieval, built by you, for flows you publish. The chat app has its own document handling that needs no assembly — drop a file into a conversation, or attach files to an agent. If what you want is "ask questions about this PDF", use Chat with a document and skip the pipeline entirely.