Built-in agent tools — and the vector store numbers

A tool is something an agent can reach for when answering is not enough. Four of them are built in, and questions here almost always work by describing a need and expecting you to name the tool.

Reading the requirement, naming the tool

What the requirement describes	The tool	Worth knowing
Answering from your own uploaded documents	File Search	Backed by a vector store. Answers are grounded and cited.
Doing maths, analysing data, producing a chart	Code Interpreter	Writes and runs code in a sandboxed container
Current information from the internet	Web Search	Leaves your compliance boundary
Calling your own code	Function	You dispatch it yourself — see the next card

The web search row connects back to network isolation. It is the tool that quietly breaks a "data stays inside the boundary" requirement.

The type strings are lowercase

Spelling is tested, so it is worth reading these carefully rather than recognising them: code_interpreter, web_search, web_search_preview, file_search, function.

Underscores, all lowercase. Note that web search has both a standard and a preview variant.

Setting up File Search

Two steps:

1. Create the vector store — a database that holds text as numeric representations so it can be searched by meaning rather than by exact words.
2. Upload the documents and poll until they are ready, using upload_and_poll.

Uploading automatically triggers a three-stage sequence: chunk, then embed, then index. Chunking splits documents into passages; embedding turns each passage into numbers; indexing makes them searchable.

The numbers

These are the kind of constants that get asked directly:

• 10,000 files per vector store
• 512 MB maximum file size, roughly 5 million tokens
• Default chunk size 800 tokens, with an overlap of 400
• Embedding model text-embedding-3-large, at 256 dimensions
• A maximum of 20 chunks placed in context

The overlap is worth a moment's thought. At 800 tokens with 400 of overlap, consecutive chunks share half their content. That redundancy exists so a passage split across a chunk boundary still appears whole somewhere.

Why tools exist at all

Three reasons, and they are a decent summary of what separates an agent from a chatbot:

• Real-time information, which beats the model's frozen training cutoff.
• The ability to act, not merely to answer.
• Enterprise grounding and workflows — reaching your systems and your data.

The trap

"Answer questions from our uploaded policy documents" is File Search.

Not web search — the documents are yours, not the internet's. And not Code Interpreter, which is the more tempting wrong answer because it sounds capable and general. Code Interpreter runs code. It does not retrieve your documents.