What vector stores do
OpenAI's retrieval guide describes vector stores as managed file collections. When you add a file, OpenAI chunks, embeds, and indexes it. You can search the store directly or make File Search available to a Responses API model.
| Need | Start with | Still verify |
|---|---|---|
| Answer one question about one document | A direct file input | The document is approved and the answer can be sent |
| Search maintained policy sources | Vector store and File Search | Source freshness, permissions, evidence, and handoff |
| Remove an outdated policy | Replace, process, then test | That searches no longer surface old content |
Adding a vector-store file is asynchronous. Use OpenAI's documented polling helpers or inspect status before you test a response. Removal is eventually consistent, so a search can return old content briefly. Plan a transition rather than treating a delete as immediate.
There are two different retrieval jobs to keep straight. A direct search is useful when your application wants the retrieved passages and will decide what to show next. File Search is a hosted tool for a Responses API request; it can retrieve from the vector store while the model prepares an answer. Neither route makes the retrieved text an approved customer promise. Your application still needs a rule for what source may be used and when a person should take over.
Treat file preparation as a release step. Keep the source file, the intended audience, and the replacement date with the change record. When a revised installation manual supersedes an old one, process the revision, test questions that used to retrieve the old wording, and only then retire the old source. This catches a common failure mode: search is technically working, but it is working from the wrong document.
Test retrieval as a support decision
Build a small evaluation set from approved questions: a normal policy question, an exception, an older wording that must not win, and a case needing customer-specific evidence. Review the retrieved passage as well as the final answer. A plausible response can still cite an obsolete source or promise an outcome that must be decided by a person.
OpenAI prices vector-store storage from parsed chunks and embeddings; model and tool use are separate. Check the current retrieval pricing before deployment. Cost does not replace quality checks: run your exact documents and questions, then review source selection and the response boundary.
For a support team, acceptance criteria should be answerable without guessing: the response identifies the current version of the guide, quotes or links the relevant prerequisite, refuses to invent a technical fix, and sends unresolved cases to the stated owner. Save the test question and retrieved evidence with the release record. That gives a reviewer something concrete to compare when the source, prompt, or model changes later.
Use eesel CLI to inspect support knowledge
OpenAI vector stores are retrieval infrastructure. eesel is a ready-to-work helpdesk teammate. Its CLI is another way to operate that teammate: a person works in a terminal, scripts consume JSON, and Claude Code, Codex, or Cursor can prepare a bounded review. The teammate and workspace are the same ones visible in the dashboard.
For example, a device-installation guide is revised after a firmware release, but customers still receive setup answers from an older source. Ask a coding agent for a read-only report: which sources are connected, whether content has downloaded, which manual files exist, and which instruction tells the teammate to cite setup prerequisites. The report must not connect accounts, change permissions, or enable automation.
With Node.js 18.17 or newer, select the teammate first:
npx @eesel/cli login
npx @eesel/cli whoami
npx @eesel/cli agents
TARGET_EESEL_AGENT="paste-agent-id-or-name-here"
npx @eesel/cli status --agent "$TARGET_EESEL_AGENT"
npx @eesel/cli integrations --agent "$TARGET_EESEL_AGENT"
npx @eesel/cli files ls --agent "$TARGET_EESEL_AGENT"
npx @eesel/cli instructions --agent "$TARGET_EESEL_AGENT"
After a person approves an updated file, preview the write before it happens:
npx @eesel/cli files upload ./approved-device-installation-guide.pdf --agent "$TARGET_EESEL_AGENT" --dry-run
Then upload it, confirm it is present, and use a fresh named chat with redacted data. Contain any action outside this test first.
npx @eesel/cli files upload ./approved-device-installation-guide.pdf --agent "$TARGET_EESEL_AGENT"
npx @eesel/cli files ls --agent "$TARGET_EESEL_AGENT"
npx @eesel/cli new --name "device-guide knowledge check" --agent "$TARGET_EESEL_AGENT"
npx @eesel/cli chat "A customer cannot complete device setup after the firmware update. Identify the required prerequisites, cite the approved guide, and state what a human must check if the steps do not resolve the issue." --agent "$TARGET_EESEL_AGENT"
Check that the reply names the current prerequisite, cites the approved guide, and makes an honest handoff when the facts are incomplete. This is an eesel support test, not a vector-store query. If it creates a held action, inspect npx @eesel/cli approvals --agent "$TARGET_EESEL_AGENT", then confirm the ticket or linked technical-system result before telling a customer it succeeded.
The eesel Activity page lists work with filters for approved, rejected, and pending items.
Choose retrieval for retrieval
Use vector stores when you are building a custom application that needs managed retrieval. Use eesel when the immediate job is operating a support teammate with approved knowledge and workflows.
Test revised support knowledge with eesel
Try eesel to inspect the same teammate from terminal or dashboard, test questions about a revised installation guide, and review outcomes before expanding automation.
Frequently asked questions
What is an OpenAI vector store?
It is a managed collection of files prepared for retrieval. When a file is added, OpenAI chunks, embeds, and indexes it so a model or a direct search can retrieve relevant passages. It is retrieval infrastructure, not proof that a support answer is current or permitted.
How do OpenAI vector stores work with File Search?
File Search can retrieve passages from a vector store for a Responses API request. A team still chooses the approved source set, access rules, prompt, customer-facing answer, and any action that follows it.
When is a vector-store file ready?
Adding a file is asynchronous. OpenAI supplies polling helpers such as uploadAndPoll and createAndPoll; do not rely on a raw upload completing before the vector-store file has finished processing.
What happens when a file is removed?
Removal is eventually consistent, so a search can briefly return removed content. Verify a replacement source is processed and re-test the intended questions before relying on the change.
How much does OpenAI vector-store storage cost?
Check OpenAI's current retrieval pricing before deployment. Storage is measured from parsed chunks and embeddings, while model and tool usage are separate costs.
Where does eesel CLI fit with vector stores?
eesel CLI does not create or query OpenAI vector stores. It operates an eesel teammate in the same workspace as the dashboard, letting a person, script, or coding agent inspect connected knowledge, review instructions, and test an approved support scenario.
