Fixes #15649

Description

The OpenAI retry decorator for both LLM and embeddings integrations currently uses a fixed exponential backoff for all retryable errors. When the server responds with a 429 status and includes a Retry-After header, the client should wait the server-specified duration instead of guessing with exponential backoff.

Without this fix, the retry loop can either wait too long (wasting time when the server says "try again in 1 second" but backoff says "wait 30 seconds") or not long enough (retrying before the rate limit window resets, burning through all retries uselessly).

Changes

New: _WaitRetryAfter wait strategy (added to both embeddings and LLM utils.py):

• Subclasses tenacity's wait_base to integrate cleanly with the existing retry stack
• On RateLimitError: extracts Retry-After from response.headers (httpx.Headers, case-insensitive)
• Caps the wait at 120 seconds to prevent a misbehaving server from stalling indefinitely
• Falls back to the existing exponential backoff for all other errors, missing headers, or unparseable values

New: _parse_retry_after helper:

• Extracts and validates the Retry-After header value
• Handles edge cases: missing response, missing headers, non-numeric values, negative values, empty strings

No breaking changes: The function signature of create_retry_decorator() is unchanged. Existing behavior is preserved for non-RateLimitError exceptions and when the Retry-After header is absent.

Files Changed

Testing

35 new unit and integration tests covering:

• Header parsing: integer, float, zero, missing, non-numeric (HTTP-date), negative, empty, case-insensitive, no response object
• Wait strategy: uses header value, caps at 120s maximum, falls back for missing header, falls back for non-RateLimitError, falls back for unparseable header, falls back when outcome is None
• Integration: full decorator stack respects Retry-After, retries exhaust at max_retries, non-RateLimitError still retries with exponential backoff

# Embeddings: 31 passed (6 existing + 17 new + 8 utils)
cd llama-index-integrations/embeddings/llama-index-embeddings-openai
pytest tests/ -v

# LLM: 41 passed (18 new + 1 existing retry + 22 existing utils)
cd llama-index-integrations/llms/llama-index-llms-openai
pytest tests/test_retry_after.py tests/test_openai_utils.py tests/test_openai.py::test_completion_model_with_retry -v

All existing tests pass unchanged.

Context

This is a follow-up to #14801 / PR #20712 (token-bucket rate limiter) which added proactive rate limiting. This PR addresses the reactive side: when a 429 does occur, the client now waits the exact amount of time the server specifies rather than guessing.

Azure OpenAI inherits from OpenAI (class AzureOpenAI(OpenAI)) so this fix applies to Azure automatically.