Non-destructive retrying on nemotron-parse API

[jamesbraza]

Previously we could have a destructive interference with retrying in nemotron-parse:

1. First retry 3X API timeouts
2. Then retry 3X bounding box errors

This PR tightens up retrying to just retry 3X cumulatively

---

Note

Medium Risk
Touches external API retry/backoff behavior, which can change throughput and error handling under load; logic is small and covered by a new unit test but still affects request timing.

Overview
Tightens Nvidia nemotron-parse retry behavior so failures across bbox validation errors, rate-limit TimeoutErrors, and LiteLLM 408 timeouts share a single stop_after_attempt(3) budget instead of potentially retrying in multiple stacked phases.

Adds a custom tenacity wait strategy that applies exponential backoff only for the rate-limit TimeoutError path (not for inference-timeout litellm.Timeout), and updates the LiteLLM timeout type check accordingly; includes a focused unit test for the new wait/backoff logic.

^{Written by Cursor Bugbot for commit 84efd89. Configure here.}

[Copilot]

Pull request overview

This PR adjusts retry behavior for the nemotron-parse Nvidia API calls to avoid “stacked” retry loops that could previously result in more than 3 total retries across different failure modes.

Changes:

• Combine multiple Tenacity retry decorators into a single retry policy capped at 3 total attempts.
• Introduce a custom Tenacity wait function to apply exponential backoff only for the TimeoutError (rate-limit) failure mode.
• Add a unit test validating the conditional backoff behavior.

Reviewed changes

Copilot reviewed 2 out of 2 changed files in this pull request and generated 1 comment.

File	Description
packages/paper-qa-nemotron/src/paperqa_nemotron/api.py	Consolidates retry logic and adds conditional exponential backoff helper.
packages/paper-qa-nemotron/tests/test_api.py	Adds test coverage for the new conditional wait/backoff logic.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.