fix: Address some issues in the split pdf logic

[awalker4]

We've encountered some bugs in the split pdf code. For one, these requests are not retried. With the new split_pdf_allow_failed=False behavior, this means one transient network error can interrupt the whole doc. We've also had some asyncio warnings such as ... was never awaited.

This PR adds retries, cleans up the logic, and gives us a much better base for the V2 client release.

Changes

Return a "dummy request" in the split BeforeRequestHook

When the BeforeRequestHook is called, we would split up the doc into N requests, issue coroutines for N-1 requests, and return the last one for the SDK to run. This adds two paths for recombining the results. Instead, the BeforeRequest can return a dummy request that will get a 200. This takes us straight to the AfterSuccessHook, which awaits all of the splits and builds the response.

Add retries to the split requests

This is a copy of the autogenerated code in retry.py, which will work for the async calls. At some point, we should be able to reuse the SDK for this so we aren't hardcoding the retry config values here. Need to work with Speakeasy on this.

Clean up error handling

When the retries fail and we do have to bubble up an error, we pass it to create_failure_response before returning to the SDK. This pops a 500 status code into the response, only so the SDK does not see a 502/503/504, and proceed to retry the entire doc.

Set a httpx timeout

Many of the failing requests right now are hi_res calls. This is because the default httpx client timeout is 5 seconds, and we immediately throw a ReadTimeout. For now, set this timeout to 10 minutes. This should be sufficient in the splitting code, where page size per request will be controlled. This is another hardcoded value that should go away once we're able to send our splits back into sdk.general.partition

Testing

Any pipelines that have failed consistently should work now. For more fine grained testing, I tend to mock up my local server to return a retryable error for specific pages, a certain number of times. In the general_partition function, I add something like

    global num_bounces # Initialize this somewhere up above
    page_num = form_params.starting_page_number or 1
    if num_bounces > 0 and page_num == 3:
        num_bounces -= 1
        logger.info(page_num)
        raise HTTPException(status_code=502, detail="BOUNCE")

Then, send a SDK request to your local server and verify that the split request for page 3 of your doc is retrying up to the number of times you want.

Also, setting the max concurrency to 15 should reproduce the issue. Choose some 50+ page pdf and try the following with the current 0.25.5 branch. It will likely fail with ServerError: {}. Then try a local pip install off this branch.

s = UnstructuredClient(
    api_key_auth="my-api-key",
)

filename = "some-large-pdf"
with open(filename, "rb") as f:
    files=shared.Files(
        content=f.read(),
        file_name=filename,
    )

req = operations.PartitionRequest(
    shared.PartitionParameters(
        files=files,
        split_pdf_page=True,
        strategy="hi_res",
        split_pdf_allow_failed=False,
        split_pdf_concurrency_level=15,
    ),
)

resp = s.general.partition(req)

if num_elements := len(resp.elements):
    print(f"Succeeded with {num_elements}")

[amanda103]

tested locally - looks good to me!