Which file did that message attach?

A customer says the answer came from the policy file, or says the assistant invented a page that was never uploaded. Do not treat the reply as something a file produced, and do not ask the model which file it used, until three proofs are in a log you stored. Name the file that message attached: the message id, the file id, the purpose, and the byte count. If the attachment list is empty, write empty, and do not invent a filename to fill the hole. Show a different file whose id, or whose purpose, is not this one. Show another thread that does not hold this message. If any of the three is a recollection instead of a stored object, you have not named the file. You have a sentence and a guess.

This is the file on one message. It is not the store. Which store did that file join? is the vector store id a file was attached to. A file can exist while no store lists it, and a store can list a file the message does not. It is not the citation. Which document did that citation name? is the document index and the span inside a reply. A citation can name a file the message's attachment list does not hold, and an attachment can be present while the citation span is empty. It is not the schema the call required. Which schema did that call require? is the schema on the request. A schema is not a byte count. It is not the coupon on a checkout, and it is not the label on an issue. Those are different objects, and they are not this id.

None of what follows is a score for this product, and none of it is legal advice. The fields below are one vendor's current description of a file and of a message, and the samples are not the same object. The counts below that are a 2024 benchmark of other people's PDFs. A byte count on an example is not a census of your documents. A paper that scored screenshots of other documents cannot name the file on the message you stored tonight. They do tell you which numbers people mix up when they try to sound precise about an upload they have not opened.

OpenAI's retrieve-file example and its list example both print the id file-abc123, and they do not print the same bytes. The retrieve example, which is also the upload response for a fine-tune file, has bytes 120000, created_at 1677610602, expires_at 1680202602, filename mydata.jsonl, purpose fine-tune, and status processed. The id is 11 characters. The filename is 12 characters. 12 is not 11. expires_at minus created_at is 2592000 seconds. 2592000 divided by 86400 is 30 exactly. The upload page says files with purpose batch expire after 30 days by default, and that expires_after.seconds must be between 3600 and 2592000. 3600 is one hour. 30 days is not one hour. This particular example lands on the maximum, not on the one-hour floor. The list example's first object is also id file-abc123, and its bytes are 175, not 120000. Its filename is salesOverview.pdf, 17 characters, and its purpose is assistants, not fine-tune. 175 is not 120000. Do not treat the shared id as one object (OpenAI, Files).

The list example's second object, and the two expiry clocks, are a third disagreement. file-abc456 has bytes 140, filename puppy.jsonl, 11 characters, purpose fine-tune, and status processed. 140 is not 175. 175 plus 140 is 315. 315 is not 120000. Both list objects print expires_at 1677614202, the same integer, on two different files. The retrieve example's expires_at is 1680202602. 1680202602 minus 1677614202 is 2588400. 2592000 minus 2588400 is 3600, so the list expiry is one hour short of a 30-day gap from the retrieve expiry, and that subtraction is between two expires_at values, not a lifetime. The list expiry is exactly 3600 seconds after the retrieve example's created_at. 1677614202 minus 1677610602 is 3600. Same id in the list, different clock. The list file's own created_at is 1613677385. 1677614202 minus 1613677385 is 63936817 seconds, which is 740 days and 817 seconds. 817 is 13 minutes and 37 seconds. 740 days is not 30 days. The second list file was created at 1613779121. The gap from the first is 101736 seconds, which is 1 day, 4 hours, 15 minutes, and 36 seconds. has_more is false. The list holds 2 objects. The list endpoint's default limit is 10000, and the maximum is 10000. 2 is not 10000.

The upload page's limits are not the message, and the purpose lists on the two sections are not the same length. An individual file can be up to 512 MB. A batch file can be up to 200 MB. 512 is not 200. A project can store up to 2.5 TB, and the page says there is no organization-wide storage limit. 2.5 TB is not 120000 bytes. Uploads are rate-limited to 1000 requests per minute. Vector-store attachment is a different limit, 2000 files per minute. 2000 divided by 1000 is 2. The page does not say the attachment limit is twice the upload limit as a design claim. It prints both integers. The Assistants note says files up to 2 million tokens. 2 million tokens is not 120000 bytes, and it is not 175 bytes. The upload purpose list is assistants, batch, fine-tune, vision, user_data, and evals. 6 names. The file object's supported purposes add assistants_output, batch_output, and fine-tune-results, and the sentence does not include evals. That list is 8 names. 6 is not 8. evals is on the request and not in that object sentence. Status is uploaded, processed, or error. 3 names. The fine-tune example is processed. The schema placeholder is uploaded. processed is not uploaded.

The message example that a person would actually read stores no file at all. Create message for thread_abc123 returns id msg_abc123, object thread.message, created_at 1713226573, completed_at 1713226573, assistant_id null, run_id null, role user, status completed, attachments an empty array, and annotations an empty array. The content value is How does AI work? Explain it in simple terms. 45 characters. 45 is not 120000. created_at and completed_at are the same integer, so the gap is 0 seconds. thread_abc123 is 13 characters. msg_abc123 is 10 characters. 10 is not 11, so the message id is not the file id even before you notice the prefix. An empty attachment list is a stored fact. It is not a missing page you are allowed to fill with file-abc123. The schema's other example, the one with placeholder strings, prints attachments with file_id file_id and a tool type code_interpreter, content type image_file, completed_at 0, created_at 0, incomplete_at 0, status in_progress, and incomplete_details.reason content_filter. Status in_progress, a completed_at of 0, and an incomplete reason, on one object. The realistic example is status completed with incomplete_at null. 0 is not 1713226573. in_progress is not completed. Do not mix the two examples because both are on the messages page (OpenAI, Messages).

The message object's other caps are a different scale from a PDF benchmark. Metadata is a set of 16 key-value pairs. Keys have a maximum length of 64 characters. Values have a maximum length of 512 characters. The schema example's key is foo, 3 characters, and its value is the word string, 6 characters. 3 is not 64. 6 is not 512. 16 is not 2 attachments. Image detail is auto, low, or high, and the default is auto. 3 names, the same count as file status, and not the same enum. Incomplete reasons are content_filter, max_tokens, run_cancelled, run_expired, and run_failed. 5 names. 5 is not 3. Message status is in_progress, incomplete, or completed. 3 names again. The realistic example is completed, and its attachment list is still empty. A completed message with no file is not a file you may infer. Role is user or assistant. 2 names. The example role is user, and assistant_id is null. A null assistant id is not a file id.

The benchmark people quote when they want a score about long documents did not open this message. Yubo Ma and colleagues, including groups at Nanyang Technological University and Shanghai AI Laboratory, present MMLongBench-Doc: 1,062 questions on 130 PDF documents, average 49.4 pages, and 20,971 textual tokens in the abstract. The introduction prints 20,970.9 tokens. The abstract's 20,971 is that figure rounded. It is not 120000 bytes, and it is not 175 bytes. The introduction says the questions are 467 single-page, 353 cross-page, and 242 unanswerable. 467 plus 353 plus 242 is 1,062. That addition matches the abstract. A later paragraph says there are 415 single-page questions, 353 cross-page, and 242 unanswerable. 415 plus 353 plus 242 is 1,010. 1,010 is not 1,062. 467 minus 415 is 52. The page does not print 52 or 1,010. 184 sourced questions plus 878 new ones is 1,062, which matches the abstract and does not match the 415 sentence. Do not pick the sentence that is easier to say (Ma et al., MMLongBench-Doc).

The same paper's rates are checkable, and they are still not a file id. 353 divided by 1,062 is 0.3324, which the abstract prints as 33.2% cross-page. 242 divided by 1,062 is 0.2279, which the abstract prints as 22.8% unanswerable. The printed rates match those two divisions. They do not match the 415 sentence, because 415 divided by 1,062 is 0.3908, and the page does not print 39.1% as the single-page share. They evaluate 14 vision-language models, 4 proprietary and 10 open-source. 4 plus 10 is 14. That addition matches. GPT-4o scores 42.7% F1. GPT-4V scores 31.4%. 42.7 minus 31.4 is 11.3. The page says the second model lags by over 10. 11.3 is over 10, and the page does not print 11.3. A footnote defines an equivalent page number using resolution 240. The experiment section says screenshots were taken at 120 DPI. 120 is not 240. Table 2 prints average pages 49.4 and a medium of 31. 31 is not 49.4. Ten annotators and a three-round review are not 14 models. None of these cells is file-abc123, and none of them is the empty attachment list on msg_abc123.

Do the three proofs on one real message, in the log that actually holds it. First, the binding. Copy the message id, the file id, the purpose, and the bytes. If attachments is an empty array, write empty. Do not infer file-abc123 because a sample file prints it, and do not infer mydata.jsonl because a sample filename prints it. If purpose is fine-tune on the file and the message was a user question, write both, because they are not the same fact. A citation inside the prose is not an attachment. Second, the other file. A different file, with a different id or a different purpose, is not this one. file-abc123 at 120000 bytes and file-abc123 at 175 bytes are that shape on the sample pages, and they are still samples. A benchmark of 130 PDFs is not that second file. If you cannot show a file that is not this one, you have a counter, not a record. Third, the other thread. A thread that belongs to someone else does not contain this message id. If the only log is a single shared table and you cannot say whose message it is, write that. A tenant you cannot name is the finding. It is not a reason to tell the customer the answer came from salesOverview.pdf.

Three moves feel like the proof and are a second problem. Pasting a filename into the reply because the attachment list was empty, and then describing the missing id as a citation. Quoting 42.7% because it looks precise, while the message you were asked about is one row, and the study that printed 42.7% scored other people's documents. Asking the model which file it used. The model will name a plausible document. It will not open the message. A store you have not kept, while the customer is arguing about which index was searched, is which store did that file join?. Opening that screen does not name the attachment on this message.

The path is ordinary when an operator can open one customer's documents without opening the next customer's documents as the way to prove it. Haystack is one search layer across documents, data lakes, and repositories. The shape that matters here is one customer's file you can point at, distinct from the next customer's file. That case is not a message-attachment ledger. Nothing in the description is a headcount, a duration, or a result figure. If the only control on the screen returns a paragraph and will not name a file id, you do not have this proof. You have a sentence.

There is a version of this afternoon where hiring the path is the wrong purchase. Nothing is in production, because the feature has not reached a customer. No model is being called. The request was not logged, so there is no message to name. A customer has already seen another customer's file. Naming an id does not close a leak. The wrong value was already a write. Or no metric was ever named, so a rescue would have nothing to move. The criteria for a system that is already in front of customers and still cannot show the file sit on the page for an AI pilot failed to reach production. That page is the symptom. It does not name this message, and it does not turn 42.7 into a byte count.

When the reply is already in front of a customer and the log cannot name the file, the missing object is specific. The message id, the file id, the purpose, and the bytes, a different file that is not that one, and a stored message rather than a recollection. An AI pilot failed to reach production engagement is for the system that is already in front of customers and cannot yet show those three: audit in the first one to two weeks, finished in 8 to 12 weeks, scoped after the audit, unpaid if the agreed metric does not move, rollout behind a human-review gate. The attachment you failed to store tonight is the kind of case the harness has to include, or the next release will cite a filename and still be unable to say which file was attached. The harness is AI evaluation and observability. An AI integration sprint is the shape where the acceptance sentence can require the file id on every stored message. If you do not yet know whether the pilot can be saved, a two-week AI readiness assessment hands you the findings either way. The free production-readiness diagnostic is the short way to see whether you are looking at a missing file, an empty attachment list, or a system that has not reached a customer. Those are the terms. They are not 120000, and they are not 42.7.

The coupon a checkout applied, and the label an issue carried, are different objects: which coupon did that checkout apply? and which label did that issue carry?. Do not record either on this page. A coupon id is not a file id. A label color is not a byte count. The run a thread started is a different object: which run did that thread start?. Do not record it here. A run id is not a file id. The message a run appended is a different object: which message did that run append?. Do not record it here. A message id is not a file id.

Before you tell them the answer came from the policy file, do the three proofs and write the ids. Name the message, the file, the purpose, and the bytes, or write that the attachment list was empty. Show a different file that is not that one. Show another thread that does not hold it. Do not ask the model which file it was, and do not cite an F1 in place of the id. A paper can tell you how a benchmark scored other PDFs, including the paragraph that does not add up to the abstract. The binding is the file you can defend.