When an OpenAI model on the Responses API works with a vector store in AI Engine, two-step questions can crash with No tool output found…. A quick fix is to ask for one thing per turn; the sections below explain why the bug happens and how raw‑API users can avoid it too.

What the Responses API expects

AssistantOne reply with function_call A and B
YouRun both, return both outputs
AnswerBoth tools used in one go

What happens with file_search on

AssistantSends only function_call A
YouAnswer A
AssistantNow sends function_call B
No tool output foundB should have come with A

AI Engine

With an OpenAI Vector Store, a model on the Responses API can use file_search to pull passages from your PDFs or Markdown and every WordPress helper you expose, such as wp_create_post or wp_upload_media.

All good, until someone asks a two‑part question that needs both tools. The model sends one function call, waits for your reply, then sends the second. Because the Responses API expects all calls and outputs in a single response, AI Engine throws:

No tool output found for function call…

Depending on the client, you may see it as “No tool output found for tool call” instead. Its mirror image, “No tool call found for function call output with call_id“, comes from the same mismatch: an output sent for a call the API no longer has pending.

Single‑tool prompts are fine; multi‑tool + vector store is what breaks.

Quick Fix

There is no switch to flip anymore: AI Engine uses the Responses API by itself for the OpenAI models that need it. Phrase questions so that only one tool fires per turn, or try another model for that chatbot. If it keeps happening, let us know.

Direct Use of the OpenAI API

The same bug appears in custom apps that combine file_search (or any vector tool) with two or more custom functions.

What breaks

Desired flow: the assistant replies once with [function_call A, function_call B]. You execute both and return both outputs.

Actual flow with file_search on: the assistant sends only function_call A. After you answer, it sends function_call B. The Responses API complains because B should have been in the first response.

Some models prefer to chain tools turn‑by‑turn, but the API spec demands they come together. Hence the clash.

Cleaner approaches

Collector loop: answer every function_call of a turn together, in one POST /responses with previous_response_id, and repeat until the assistant stops emitting new ones.

Selective tool set: answer the vector‑store part first, then send a second prompt for the remaining tools.

// pseudo-code (Responses API): answer every call of a turn together
let res = await openai.responses.create({ model, tools, input });
let calls = res.output.filter(item => item.type === 'function_call');
while (calls.length) {
  const outputs = await Promise.all(calls.map(async call => ({
    type: 'function_call_output',
    call_id: call.call_id,
    output: await run(call.name, JSON.parse(call.arguments))
  })));
  res = await openai.responses.create({ model, tools, previous_response_id: res.id, input: outputs });
  calls = res.output.filter(item => item.type === 'function_call');
}

For now: keep it single‑tool or single‑turn

If you see No tool output found… in your logs, ask yourself: ā€œAm I calling two tools and file_search in the same turn?ā€ If so, split the steps or use the collector loop above until the next model update.

That should keep your chats humming, whether you run WordPress with AI Engine or wire straight into the OpenAI API. Happy building!