[Fix] broaden RAG ingestion credential cleanup to AWS endpoint/identity fields

[yuneng-berri]

Relevant issues

Summary

Extends the cleanup logic in BaseRAGIngestion._load_credentials_from_config()
so caller-supplied aws_sts_endpoint and aws_web_identity_token are
removed when a litellm_credential_name is referenced and the stored
credential does not itself define those fields. Mirrors the existing
api_base handling.

Fix

The credential-overwrite loop already replaces caller-supplied entries on
keys that the stored credential defines, but the post-loop strip covered
only api_base. Generalize the strip to a tuple of three keys
(api_base, aws_sts_endpoint, aws_web_identity_token) so the resolved
vector_store_config cannot pair the credential's keys with caller-chosen
endpoint or identity values when the stored credential omits them.

Testing

• uv run pytest tests/test_litellm/proxy/rag_endpoints/test_rag_endpoints.py tests/test_litellm/vector_stores/test_vector_store_registry.py tests/test_litellm/proxy/vector_store_endpoints/test_vector_store_endpoints.py -v — 51 passed.
• Manual API verification against the proxy with three scenarios:
  1. Stored credential + caller-supplied aws_sts_endpoint to a localhost listener → after the fix, the listener receives no traffic and the call routes to the real AWS STS.
  2. Stored credential, no malicious overrides → unchanged before/after.
  3. Caller-direct keys (no stored credential) + caller-supplied aws_sts_endpoint → unchanged before/after; caller remains in control of their own endpoint.

Type

🐛 Bug Fix

Screenshots

[CLAassistant]

All committers have signed the CLA.

[veria-ai]

Low: No security issues found

This PR adds Azure GPT-5.5 model pricing entries and broadens the RAG ingestion credential cleanup to also strip caller-supplied aws_sts_endpoint and aws_web_identity_token when a stored credential is resolved. The credential cleanup change is security-positive, preventing stale or attacker-injected AWS identity fields from overriding the stored credential definition.

---

Status: 0 open
Risk: 1/10

Posted by Veria AI · 2026-04-25T21:49:26.808Z

[greptile-apps]

Greptile Summary

This PR has two independent parts: (1) a security fix in BaseRAGIngestion._load_credentials_from_config() that extends the post-credential-overwrite cleanup to strip caller-supplied aws_sts_endpoint and aws_web_identity_token when the stored credential doesn't define them (preventing SSRF via a rogue STS endpoint), and (2) new model_prices_and_context_window.json entries for Azure GPT-5.5 and GPT-5.5-pro (base and dated variants) with matching pricing and reasoning-effort tests. The credential-cleanup logic is correct and well-tested across all three documented scenarios.

Confidence Score: 5/5

Safe to merge; the credential-cleanup fix is correct and the only inline finding is a P2 test-coverage gap.

All P0/P1-level concerns from the prior review thread (missing reasoning-effort flags on dated variants) are pre-existing and already tracked. The new code in base_ingestion.py is logically sound and covers the three documented scenarios. The sole new finding is a P2 missing assertion for supports_low_reasoning_effort in the parametrized test, which does not affect production correctness.

model_prices_and_context_window.json — dated variant entries (azure/gpt-5.5-2026-04-23, azure/gpt-5.5-pro-2026-04-23) still lack reasoning-effort flags (tracked in prior review thread).

Important Files Changed

Filename	Overview
litellm/rag/ingestion/base_ingestion.py	Core security fix: generalises the post-credential-overwrite strip from api_base-only to also cover aws_sts_endpoint and aws_web_identity_token, preventing a caller from pairing a stored credential with attacker-controlled STS/identity endpoints.
model_prices_and_context_window.json	Adds Azure GPT-5.5 / GPT-5.5-pro (and their 2026-04-23 dated variants). Dated variants are missing all reasoning-effort flags that their base counterparts define (flagged in a prior review thread).
tests/test_litellm/litellm_core_utils/llm_cost_calc/test_llm_cost_calc_utils.py	Adds pricing and reasoning-effort tests for Azure GPT-5.5 entries; reasoning-effort parametrize doesn't cover supports_low_reasoning_effort or dated variants.
tests/test_litellm/test_utils.py	Adds supports_low_reasoning_effort to the JSON-schema validator — a necessary companion to the new azure/gpt-5.5-pro entry that first uses this field.

Flowchart

%%{init: {'theme': 'neutral'}}%%
flowchart TD
    A[__init__: vector_store_config\nloaded from caller ingest_options] --> B{litellm_credential_name\npresent AND credential_list set?}
    B -- No --> Z[Use caller config as-is]
    B -- Yes --> C[CredentialAccessor.get_credential_values]
    C --> D{credential_values\nempty?}
    D -- Yes --> Z
    D -- No --> E[Write all credential values\ninto vector_store_config\noverwriting caller keys]
    E --> F{For each key in\napi_base / aws_sts_endpoint /\naws_web_identity_token}
    F --> G{Key in vector_store_config\nAND key NOT in credential_values?}
    G -- Yes --> H[Delete key\nremoves caller-supplied\nendpoint / identity value]
    G -- No --> I[Keep value\nstored credential owns it]
    H --> J[Resolved vector_store_config\nfully controlled by credential]
    I --> J

Loading

_{Reviews (2): Last reviewed commit: "style: black formatting" | Re-trigger Greptile}

[greptile-apps]

Dated variants missing reasoning-effort flags

azure/gpt-5.5-2026-04-23 and azure/gpt-5.5-pro-2026-04-23 omit all supports_*_reasoning_effort flags that their base counterparts define. Any caller querying these dated model keys via get_model_info will get None for those capabilities instead of the correct values, potentially causing incorrect request building (e.g., no guard against unsupported effort levels).

azure/gpt-5.5-2026-04-23 should include:

"supports_none_reasoning_effort": true,
"supports_xhigh_reasoning_effort": true,
"supports_minimal_reasoning_effort": false

azure/gpt-5.5-pro-2026-04-23 should include:

"supports_none_reasoning_effort": false,
"supports_xhigh_reasoning_effort": true,
"supports_minimal_reasoning_effort": false,
"supports_low_reasoning_effort": false

[codecov]

Codecov Report

❌ Patch coverage is 0% with 3 lines in your changes missing coverage. Please review.

Files with missing lines	Patch %	Lines
litellm/rag/ingestion/base_ingestion.py	0.00%	3 Missing ⚠️

📢 Thoughts on this report? Let us know!