LiteLLM

LiteLLM is a candidate self-hosted or managed gateway layer for normalizing multiple provider APIs, routing, fallback, spend, and policy. The update needs release-pinned implementation evidence and must not convert its compatibility table or marketing count into a universal guarantee.

provider mapping / gateway / implementation / operability / cost / governance / security

What does the author-maintained LiteLLM documentation/repository state about the SDK/proxy boundary, multi-provider normalization, routing/fallback, and gateway controls, and what remains unverified?

[S1] SOURCE_CLASS: author_maintained_repo | EXACT_TITLE: BerriAI/litellm - LiteLLM repository | AUTHOR_ORG: BerriAI / LiteLLM maintainers | CANONICAL_URL: https://github.com/BerriAI/litellm | PUBLICATION_UPDATE_ACCESS: repository accessed 2026-09-14 00:16 JST | VERSION_RELEASE_TAG: default branch; not pinned; repository advertises stable Docker tag separately | SUPPORTING_LOCATOR: repository getting-started lines 540-549 for Proxy Server versus Python SDK, unified interface, gateway features, Router retry/fallback, and stable-tag note; About lines 750-754 for gateway description | EVIDENCE_LABEL: OSS_PRACTICE + PRODUCT_VOLATILE | RELATION: support for documented implementation surface; limit for mutable branch and provider parity. [S2] SOURCE_CLASS: official_project_docs | EXACT_TITLE: LiteLLM Documentation - Getting Started / Proxy | AUTHOR_ORG: LiteLLM project | CANONICAL_URL: https://docs.litellm.ai/docs/ | PUBLICATION_UPDATE_ACCESS: current documentation accessed 2026-09-14 00:16 JST | VERSION_RELEASE_TAG: current docs; page exposes provider/model examples and config paths but no immutable release pin in this capture | SUPPORTING_LOCATOR: documentation index and proxy guidance for OpenAI-compatible gateway, provider translation, authentication, logging, cost tracking, rate limiting, and guardrails | EVIDENCE_LABEL: VERIFIED_SPEC + PRODUCT_VOLATILE | RELATION: support/context; limit because documented option is not proof of configured deployment. [S3] SOURCE_CLASS: author_maintained_repo | EXACT_TITLE: LiteLLM provider support matrix | AUTHOR_ORG: BerriAI/LiteLLM maintainers | CANONICAL_URL: https://github.com/BerriAI/litellm | PUBLICATION_UPDATE_ACCESS: matrix in repository accessed 2026-09-14 00:16 JST | VERSION_RELEASE_TAG: default branch; support cells are mutable | SUPPORTING_LOCATOR: provider/model capability matrix around repository lines 519-536; treat cells as implementation claims requiring version/provider recheck | EVIDENCE_LABEL: PRODUCT_VOLATILE + OSS_PRACTICE | RELATION: support for existence of a maintained mapping surface; limit/counterevidence to universal compatibility.

• ESTABLISHED [S1+S2 support]: LiteLLM presents two boundaries: a Python SDK for direct application integration and a Proxy Server described as a central LLM gateway. The repository says both expose a unified interface across multiple LLMs.
• ESTABLISHED [S1 support]: The repository documents gateway features including authentication/authorization hooks, multi-tenant spend tracking, logging, guardrails, caching, virtual keys, and a Router with retry/fallback across deployments.
• ESTABLISHED [S1+S3 limit]: Provider/model support is a mutable implementation matrix. A checked cell is not proof that every endpoint, parameter, modality, tool schema, error, streaming mode, or release behaves identically.
• INFERENCE [S1+S2]: LiteLLM can serve as a policy/normalization control plane, but the normalization boundary may hide provider-specific semantics and does not itself make model output, tool calls, data handling, or fallback outcomes equivalent.

The repository's 100+ provider/API count and feature descriptions are project claims, not independent benchmark evidence. Default branch, stable tags, provider adapters, and upstream APIs change. Retry/fallback can duplicate or alter requests and may change model semantics; compatible request shape does not prove compatible safety, cost, context, tool, or response behavior. No release tag, deployment, or provider API was executed in this packet.

LiteLLM release/tag/container image; Proxy versus SDK; provider adapter/version; model ID and API revision; endpoint/streaming/tool/structured-output support; credentials and secret store; auth/RBAC/virtual-key configuration; routing/fallback/retry policy; spend/rate limits; logging/guardrails/caching; data retention/residency; network egress; license/support/managed contract; audit export.

Evidence-only desk research. No LiteLLM process, proxy, model provider, key, request, or callback was run.

No cross-provider compatibility matrix was independently reproduced; no latency, cost, fallback-duplication, tool-call, schema, or safety benchmark was run.

Treat gateway logs, prompts, outputs, API keys, virtual keys, and provider metadata as sensitive. Audit should retain model/provider selected, all attempts, policy/rate decisions, request correlation, redaction status, token/cost source, and downstream result. Calculate cost at accepted outcome level, not only gateway token estimate. Test timeouts, retry idempotency, provider outage, partial stream, and cache leakage; not measured.

「LiteLLMは、Python SDKとProxy Serverを通じて複数LLM providerを共通インターフェースへ寄せ、routing／retry／fallback、認証、ログ、guardrail、spend管理等を提供すると公式リポジトリで説明している。provider互換性、品質、価格、データ処理、fallback安全性はrelease・model・設定単位で再確認する。」

Future update Work should select one pinned LiteLLM release and two providers, then run a safe read-only TEST_PROFILE for request/response, streaming, tool/schema, timeout/retry, fallback duplication, rate limit, cost attribution, redaction, audit, and provider data policy. Preserve provider-specific failure records rather than collapsing them into the gateway claim.

current stable release/commit; adapter parity; provider-specific parameter loss; tool/structured-output semantics; retry/fallback duplicate side effects; cache isolation; RBAC/audit/data-residency gates; cost calculation accuracy; public incidents; managed versus OSS feature difference.

Do not use LiteLLM's provider count as interoperability proof. Do not state that OpenAI-compatible input/output means equal model behavior, safety, tools, pricing, or SLA across providers. UNIT_CLOSE_STATUS: INDIVIDUAL_BOUNDED_CLOSE / EVIDENCE_RECORDED / UNKNOWN_PRESERVED

---