fix: support Responses input in Redis semantic cache

[balcsida]

Relevant issues

Fixes #28272

Linear ticket

N/A

Pre-Submission checklist

• I have added testing in the tests/test_litellm directory
• My PR passes all unit tests on make test-unit
• My PR scope is isolated and solves one specific problem
• I have requested a Greptile review

Screenshots / Proof of Fix

Before this change, RedisSemanticCache only extracted prompts from messages. Responses API calls pass input instead, so RedisVL store/check were skipped and the RediSearch index stayed empty.

Validation run locally:

uv run pytest tests/test_litellm/caching/test_redis_semantic_cache.py -q
...................                                                      [100%]
19 passed in 0.27s

uv run ruff check litellm/caching/caching.py litellm/caching/redis_semantic_cache.py tests/test_litellm/caching/test_redis_semantic_cache.py
All checks passed!

uv run black --check litellm/caching/caching.py litellm/caching/redis_semantic_cache.py tests/test_litellm/caching/test_redis_semantic_cache.py
All done!
3 files would be left unchanged.

I also attempted the broader caching test directory, but the local environment is missing optional dependencies for unrelated Azure/S3 cache tests (azure, boto3).

Type

Bug Fix
Test

Changes

• Added RedisSemanticCache prompt extraction for Responses API input, while preserving existing chat messages extraction.
• Supported string input and structured Responses input content lists.
• Updated sync cache lookup to pass original kwargs through to backend caches so input and metadata are available to semantic caches.
• Added regression tests for sync set/get, async set/get, structured Responses input extraction, and wrapper passthrough.

[CLAassistant]

Thank you for your submission! We really appreciate it. Like many open source projects, we ask that you all sign our Contributor License Agreement before we can accept your contribution.
5 out of 7 committers have signed the CLA.

✅ TorvaldUtne
✅ mubashir1osmani
✅ balcsida
✅ ro31337
✅ cwang-otto
❌ oss-agent-shin
❌ IshaMeera
_{You have signed the CLA already but the status is still pending? Let us recheck it.}

[codecov]

Codecov Report

❌ Patch coverage is 88.88889% with 8 lines in your changes missing coverage. Please review.

Files with missing lines	Patch %	Lines
litellm/caching/redis_semantic_cache.py	90.00%	7 Missing ⚠️
litellm/caching/caching.py	50.00%	1 Missing ⚠️

📢 Thoughts on this report? Let us know!

[greptile-apps]

Greptile Summary

This PR fixes RedisSemanticCache ignoring Responses API calls by extracting a cacheable prompt from the input field when messages is absent, and updates the sync Cache.get_cache to forward full **kwargs to backend caches (making it consistent with the async path that already did this).

• _get_prompt_from_kwargs / _collect_responses_input_text / _coerce_response_input_value are added to RedisSemanticCache to handle string input, nested structured content lists, and Pydantic model objects.
• The sync Cache.get_cache now passes **kwargs rather than messages=messages to the backend, enabling semantic caches to access input and other Responses API fields.
• Eight new unit tests cover sync/async set+get with string input, structured Responses input flattening, edge-case prompt extraction, and the Cache wrapper passthrough; all are fully mocked.

Confidence Score: 4/5

Safe to merge; the fix is well-scoped and the async path already behaved this way, so the sync change is a consistency fix rather than new ground.

Both the dict and object branches of _collect_responses_input_text return early when any text-keyed attribute is a present-but-blank string, so a fallback key with real content would be silently missed. This is an edge case unlikely to surface with real Responses API payloads.

litellm/caching/redis_semantic_cache.py — the text-key fallback loop in both the dict and object branches of _collect_responses_input_text.

Important Files Changed

Filename	Overview
litellm/caching/redis_semantic_cache.py	Adds _get_prompt_from_kwargs, _collect_responses_input_text, and _coerce_response_input_value to extract prompts from Responses API input alongside existing messages support; replaces four hard-coded messages extractions in set/get (sync+async) with the new helper. Minor edge case in text-key fallback logic.
litellm/caching/caching.py	Removes local messages extraction in sync get_cache and passes full **kwargs to backend cache, making it consistent with the already-existing async path behavior.
tests/test_litellm/caching/test_redis_semantic_cache.py	Adds 8 new unit tests covering sync/async set+get with string input, structured Responses input flattening, prompt extraction edge cases (model objects, None, whitespace), and the Cache wrapper passthrough. All mocked, no real network calls.
tests/test_litellm/interactions/test_openapi_compliance.py	Adds "budget_exceeded" to the expected Responses API status enum in the OpenAPI compliance assertion.

_{Reviews (1): Last reviewed commit: "test: include budget exceeded interactio..." | Re-trigger Greptile}

[oss-pr-review-agent-shin]

🤖 litellm-agent: Auto-merge skipped — the staging branch shin_agent_oss_staging_05_19_2026 has 5 commit(s) not in your branch. Merging as-is would produce a confusing diff on the staging PR.

Please rebase your branch onto shin_agent_oss_staging_05_19_2026 and push; the agent will re-review automatically.

[balcsida]

May I ask why this was closed?

[Sameerlite]

@balcsida The branch was deleted when cleaning old branches up. Can you use the latest internal staging as base and recreate the PR? sorry for the trouble.

[balcsida]

Fair enough, @Sameerlite
Unless there was another constraint I’m missing, updating the base branch seems like it would have been the cleaner option than closing the PR.