Assistants

An assistant is the part of your AI setup that decides how the AI answers: what it is told to do, which model serves it, which tools it may use, and which of your documents it may read from.

Assistants live in the Control Room, under your OpenAI integration. They belong to that integration and to your organization, and one assistant can be used by as many bots as you like — change it once and every bot behind it answers differently.

What changed in August 2026#

OpenAI permanently removed its Assistants API on 26 August 2026, and the dashboard where assistants used to be created went with it. Assistants are now created and edited in the Control Room, and Quriobot talks to OpenAI through its Responses API instead.

Nothing about this changed what your bots do. Your OpenAI account, your API key, your vector stores and the documents in them were not affected.

Where to find them#

Integrations → your OpenAI integration → Assistants.

The list shows every assistant on that integration: its name, the model that answers, and when it was last changed. Open one to edit it, or use Create to add a new one. You can also create one without leaving the bot you are building, straight from a Chat Completion variable.

Instructions#

The instructions are the assistant’s system prompt — sent ahead of every conversation, before anything the visitor says. They decide most of how the assistant behaves: what it is, what it should and should not talk about, what tone it takes, what to do when it does not know something.

Write them as you would brief a new colleague on their first day. Be specific about the boundaries: an assistant told only “be helpful” will cheerfully answer questions about anything at all.

The model#

Pick the model that answers. The list is read live from your own OpenAI account, so it always reflects the models you actually have access to — including ones released after this article was written. The newest capable model is offered first.

The model you pick here is the model used. Nothing substitutes it silently.

Tools#

Tools are what the assistant may do beyond answering from its instructions and its knowledge. Each one is billed to your own OpenAI account.

• Web search — the assistant may look something up on the public web and answer from what it finds. Sources are shown with the answer, so a visitor can see where a claim came from. Useful when the answer depends on anything that changes: prices, opening hours, availability, news.
• Code interpreter — the assistant may run code to work something out: a calculation, a chart, a transformation of a file the visitor uploaded. Anything it draws is shown in the reply.
• Image generation — the assistant may draw a picture and show it in the reply: a mock-up, a diagram, an illustration. Each image takes several seconds.
• Function tools — the assistant may ask your bot to run a function and use the result. This is how an assistant reaches your own systems: a stock lookup, a booking, a customer record. See Function calling in the Chat Completion article for the script that answers the call.

File search is not in this list because it switches itself on: attach a storage under Knowledge and the assistant can search it.

Knowledge#

Knowledge is the set of documents the assistant answers from. Documents live in storages — vector stores in your own OpenAI account. Quriobot records which storages an assistant is attached to; the storages and their files remain yours.

Attaching a storage#

Under Knowledge, search your storages and pick one. The picker lists every storage on the OpenAI account behind this integration, with the number of files in each, so you can attach one you already built for something else.

An assistant may have up to 5 storages attached — that is OpenAI’s own ceiling on how many one file search may cover.

Creating a storage#

Create storage makes a new, empty one on your OpenAI account and attaches it in the same step. Give it a name you will recognise later; you will be picking it from a list.

The documents in a storage#

Documents opens what a storage holds. Each file is listed with its indexing status — a document is only searchable once it shows Indexed, which takes a moment after upload for anything large.

Add documents from the same place. Accepted are text files, PDF, Word and PowerPoint documents, JSON, Markdown and source files, up to 536.9 MB each, and up to 20 files per assistant.

Removing a document takes it out of the storage and deletes it from your OpenAI account. You are asked to confirm before that happens.

Detaching versus deleting#

Two different things, deliberately kept apart:

• Detach takes the storage off this assistant. The storage and every document in it survive, and you can attach it again — to this assistant or another one.
• Delete destroys the storage and its documents on your OpenAI account. It cannot be undone, and if another assistant was using that storage, it loses those documents too.

Trying an assistant before a bot uses it#

Every assistant has a Sandbox — a conversation that runs exactly the path a live bot runs, without touching a bot and without a visitor seeing anything.

For each answer it shows what actually happened: which tools fired, what the file search matched and from which document, what a function was asked for and what it returned, which model served the request, how many tokens it cost and how long it took. When an answer disappoints, this is where you find out why, instead of guessing.

Conversations are kept for 30 days after they were last used, so you can change the instructions and compare the same question before and after.

Runs in the sandbox use your own OpenAI credits, exactly as a live conversation does.

Who can change an assistant#

Everyone in your organization who can see an integration can read its assistants and open the sandbox to see what was tried. Changing an assistant — its instructions, model, tools or knowledge — and running a test are for organization administrators.

That split exists because an assistant is shared: editing one changes what every bot behind it says to every visitor, which is not the same kind of act as reading it. If the controls are missing where you expect them, ask an administrator of your organization for access.

Assistants from before August 2026#

Assistants that existed before the shutdown were migrated for you, together with their storages and documents. Two things could not be carried across, because OpenAI deleted them along with the assistants themselves:

The instructions. Where a bot had past conversations, we read them and drafted replacement instructions from them, so the assistant behaves close to how it did before. Those are marked Review in the list and carry a notice on the assistant itself. They are our reconstruction, not your original text — read them and correct anything that does not match what you had. The marker disappears once you save your own edit.

Function tool schemas. If an assistant called a function, its JSON Schema lived on the OpenAI assistant object and is gone. Until you declare it again under Function tools, the assistant answers without ever calling that function. The bot variable’s function script is unchanged — only the declaration moved.

If you ever pressed Create prompt on an assistant in your own OpenAI dashboard before the shutdown, your original instruction text may still be there under Prompts. That listing is per project, so check each of your projects before concluding there is nothing to recover.

Where assistants are used#

• Chat Completion bot variable — the assistant answers inside a bot conversation. See Complete chat with OpenAI.
• Live chat AI Draft — the assistant drafts a reply for a human agent to send or edit. See AI-powered message drafts.
• Integration automations — a rule can keep an assistant’s knowledge up to date on a schedule, for example from a crawled website or a Google Drive folder.

Edit this page