* feat: add litellm.compress() for BM25-based context compression
Adds a compress() utility that reduces context size for LLM calls using
BM25 relevance scoring (with optional semantic embeddings via
litellm.embedding()). Messages below a token threshold pass through
unchanged; messages above are scored, ranked, and the lowest-relevance
ones replaced with stubs. Originals are cached and a retrieval tool is
injected so the model can recover dropped content on demand.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
* fix(compress): truncate high-scoring messages instead of fully stubbing them
When a relevant message was too large to fit in the token budget it was
replaced with a stub, leaving the LLM with no real content to work with.
Now the highest-scoring overflow message is truncated (first 70% + last 30%
of words) to fill the remaining budget, so the LLM always receives actual
content rather than just a retrieval pointer.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
* fix(bm25): add prefix expansion so query terms match inflected doc tokens
"cook" now matches "cooking", "auth" matches "authentication", etc.
Without this, short query terms scored 0 against longer inflected forms
in documents, causing the wrong message to be kept.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
* test: add routing correctness test and eval harness for litellm.compress()
- test_simple_compression: parametrized test verifying BM25 routes the
right message based on query ("How to cook?" keeps cooking, "Fix auth"
keeps auth content)
- eval_compression.py: end-to-end eval harness comparing baseline vs
compressed model performance on HumanEval-style coding problems
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
* feat(eval): add SWE-bench Lite compression eval harness
Uses princeton-nlp/SWE-bench_Lite_bm25_27K which bundles ~27k tokens of
BM25-retrieved repo context per problem — large enough to meaningfully
stress litellm.compress() without Docker or GitHub API calls.
Proxy eval metrics (no test runner needed):
- has_diff: model produced a valid unified diff
- file_overlap: fraction of gold-patch files in generated patch
- exact_file_match: generated patch touches exactly the right files
Run: python tests/eval_swe_bench.py --model gpt-4o --problems 10
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
* fix(eval): robust dataset loading + sys.path fix for worktree imports
- Add HuggingFace API fallback so the SWE-bench loader doesn't need
the `datasets` library (avoids pyarrow/numpy binary compat issues)
- Insert repo root into sys.path so compression module resolves
from worktrees
- Use direct import of litellm_compress to avoid __getattr__ issues
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
* improve compression quality: line-based truncation, multi-message budget, 70% default target
- Switch truncate_message from word-based to line-based splitting to
preserve code structure (function boundaries, indentation)
- Allow multiple messages to be truncated instead of burning entire
budget on one overflow message
- Raise default compression target from 50% to 70% of trigger for
better quality/cost tradeoff
- Add --compression-target CLI arg to SWE-bench eval harness
- Move tests to canonical locations (tests/test_litellm/, scripts/)
- Add docs page and sidebar entries for compress()
Eval results (5 problems, Opus, trigger=10k):
Hunk overlap delta improved from -0.417 to -0.221
Content similarity now matches baseline (+0.006)
Cost savings: 72%
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
* docs: add SWE-bench performance results to compress() docs
Include benchmark table from Opus eval (5 problems, trigger=10k)
showing 72% cost savings with file-level quality fully preserved.
Add metric explanations and eval runner examples.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
* fix(eval): use tolerance-based hunk overlap metric
The exact line-number matching was too brittle — LLM-generated patches
often target the right code region but with slightly offset line numbers.
Switch to hunk-level overlap with a 10-line tolerance window so nearby
edits count as matches. This better reflects actual patch quality.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
* feat: add compression_interception callback for LiteLLM Proxy
Add a proxy callback that automatically compresses incoming /v1/messages
payloads above a configurable token threshold, runs the retrieval tool
loop server-side, and returns the final response. This brings compress()
support to proxy deployments (e.g. Claude Code via /v1/messages).
- New callback: litellm/integrations/compression_interception/
- Proxy config: compression_interception_params in litellm_settings
- Support for input_type param in compress() (openai vs anthropic)
- Docs: proxy setup instructions with YAML config example
- Tests: 139-line unit test suite for the interception handler
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
* Revert "feat: add compression_interception callback for LiteLLM Proxy"
This reverts commit 72bd5cb152ca1df07f14a14e14a2816e188874a8.
---------
Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>
The blog CSS selectors for dark mode used descendant selectors like
[data-theme='dark'] .blog-wrapper which never matched because both
data-theme and .blog-wrapper are applied to the same <html> element
by Docusaurus. Fixed by using compound selectors (no space):
[data-theme='dark'].blog-wrapper.
Also added missing dark-mode overrides for:
- pre/code blocks in blog posts
- link colors in blog posts
- marquee items, separators, and labels on blog list page
- pagination links on blog list page
- meta text and author separators on blog list page
Co-authored-by: Cursor Agent <cursoragent@cursor.com>
Co-authored-by: Krrish Dholakia <krrish-berri-2@users.noreply.github.com>
- Add LITELLM_OIDC_ALLOWED_CREDENTIAL_DIRS to the environment variables
reference so the documentation test passes.
- Annotate the values variable in _reject_os_environ_references so it
accepts both dict.values() and list iterables.
- Log a warning when dropping callback params that carry os.environ/
references so operators notice the misconfiguration.
- Require absolute paths in oidc/file/ and correct the documented
example to use the leading-slash form.
- Drop the unused return value from _reject_os_environ_references.
- Reject os.environ/ references supplied via /health/test_connection
request params instead of resolving them; config-sourced values are
already resolved before reaching the endpoint.
- Skip os.environ/ references in dynamic callback params loaded from
per-request metadata.
- Constrain oidc/file/ to an allowed credential directory allowlist
(defaults to /var/run/secrets and /run/secrets, overridable via
LITELLM_OIDC_ALLOWED_CREDENTIAL_DIRS).
* feat(guardrails): optional skip system message in unified guardrail inputs
Made-with: Cursor
* feat(dashboard): skip_system_message_in_guardrail in guardrail UI
Add a tri-state control (inherit / yes / no) when creating or editing
guardrails so admins can set litellm_params.skip_system_message_in_guardrail
without YAML. Table edit merges existing litellm_params before PUT to avoid
wiping content-filter and other provider fields.
Document the dashboard flow in the guardrails quick start with a screenshot.
Made-with: Cursor
* fix(guardrails): type structured_messages as AllMessageValues for mypy
Use AllMessageValues in openai_messages_without_system and cast adapter
request messages so GenericGuardrailAPIInputs matches TypedDict.
Made-with: Cursor
* build: migrate packaging metadata to uv
* ci: move automation and local tooling to uv
* docker: migrate image builds and runtime setup to uv
* docs: update install and deployment guidance for uv
* chore: align auxiliary scripts and tests with uv
* test: harden test_litellm isolation
* fix: keep release and health check images self-contained
* build: pin uv tooling and health check deps
* test: isolate bedrock image request formatting from suite state
* test: cover sandbox executor requirements flow
* ci: fix circleci no-op command steps
* ci: fix circleci publish workflow parsing
* fix: stabilize remaining uv migration CI checks
* ci: increase matrix test timeout headroom
* fix: restore published docker and license coverage
* fix: restore proxy runtime build parity
* fix: restore proxy extras parity and venv migrations
* ci: persist uv path across circleci steps
* fix: keep psycopg binary in default test env
* docker: preserve prisma cache across stages
* test: run local proxy checks through uv python
* build: restore runtime deps moved into ci
* build: refresh uv lock after upstream merge
* fix: restore module import in test_check_migration after merge
The conflict resolution imported only the function but the test body
references check_migration as a module throughout.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
* fix: revert dependency promotions, remove nodejs-wheel-binaries, fix Docker layer caching
- Move google-generativeai, Pillow, tenacity back to ci group (they are
lazily imported and bloat the base SDK install needlessly)
- Remove nodejs-wheel-binaries from extra_proxy and proxy-dev (redundant
in Docker where system Node.js is already installed via apk)
- Remove all nodejs-wheel node replacement and venv npm patching blocks
from Dockerfiles since the wheel is no longer installed
- Add --no-default-groups to CodSpeed benchmark workflow so the benchmark
environment matches the old minimal pip install footprint
- Apply standard uv two-phase Docker pattern: copy metadata first, install
deps (cached layer), then copy source and install project
- Replace CircleCI enterprise no-op with proper uv sync command
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
* chore: regenerate uv.lock after removing nodejs-wheel-binaries
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
* fix(ci): use cache/restore instead of cache to prevent cache poisoning
The old workflow used actions/cache/restore (read-only). The uv migration
changed it to actions/cache (read-write), which zizmor flags as a cache
poisoning risk. Restore the safer read-only variant.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
* fix(ci): disable setup-uv built-in cache to silence cache-poisoning alert
The setup-uv action enables caching by default, which zizmor flags as a
cache poisoning risk. Disable it since we already use a read-only
cache/restore step.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
* fix(ci): disable setup-uv cache in publish workflow
Silences zizmor cache-poisoning alert. Publishing workflow runs
infrequently on protected branches so caching adds no real benefit.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
* fix(test): remove duplicate verbose_logger mock in test_check_migration
The logger was patched twice — first via mocker.patch() then via
mocker.patch.object(autospec=True). The second call fails because
autospec cannot inspect an already-mocked attribute. Remove the
redundant first patch.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
* fix(ci): free disk space before Docker build in test-server-root-path
The Dockerfile.non_root build ran out of disk on the CI runner. Remove
Android SDK, .NET, Boost, and GHC toolchains (~12GB) to free space.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
* Add PromptGuard guardrail integration
Add PromptGuard as a first-class guardrail vendor in LiteLLM's proxy,
supporting prompt injection detection, PII redaction, topic filtering,
entity blocklists, and hallucination detection via PromptGuard's
/api/v1/guard API endpoint.
Backend:
- Add PROMPTGUARD to SupportedGuardrailIntegrations enum
- Implement PromptGuardGuardrail (CustomGuardrail subclass) with
apply_guardrail handling allow/block/redact decisions
- Add Pydantic config model with api_key, api_base, ui_friendly_name
- Auto-discovered via guardrail_hooks/promptguard/__init__.py registries
Frontend:
- Add PromptGuard partner card to Guardrail Garden with eval scores
- Add preset configuration for quick setup
- Add logo to guardrailLogoMap
Tests:
- 30 unit tests covering configuration, allow/block/redact actions,
request payload construction, error handling, config model, and
registry wiring
* Fix redact path and init ordering per review feedback
- P1: Update structured_messages (not just texts) when PromptGuard
returns a redact decision, so PII redaction is effective for the
primary LLM message path
- P2: Validate credentials before allocating the HTTPX client so
resources aren't acquired if PromptGuardMissingCredentials is raised
- Add tests for structured_messages redaction and texts-only redaction
* Harden PromptGuard integration: fail-open, event hooks, images, docs
- Add block_on_error config (default fail-closed, configurable fail-open)
- Declare supported_event_hooks (pre_call, post_call) like other vendors
- Forward images from GenericGuardrailAPIInputs to PromptGuard API
- Wrap API call in try/except for resilient error handling
- Add comprehensive documentation page with config examples
- Register docs page in sidebar alongside other guardrail providers
- Expand test suite from 32 to 40 tests covering new functionality
* Fix dict[str, Any] -> Dict[str, Any] for Python 3.8 compat
* Address remaining Greptile feedback: timeout, redact guard
- Add explicit 10s timeout to async_handler.post() to prevent
indefinite hangs when PromptGuard API is unresponsive
- Guard redact path: only update inputs["texts"] when the key
was originally present, avoiding phantom key injection
- Add test: redact with structured_messages only does not create
texts key (41 tests total)
* Fix CI lint: black formatting, add PromptGuardConfigModel to LitellmParams
- Reformat promptguard.py to match CI black version (parenthesization)
- Add PromptGuardConfigModel as base class of LitellmParams for proper
Pydantic schema validation, consistent with all other guardrail vendors
- Use litellm_params.block_on_error directly (now a typed field)
* Address Greptile review: redact path, null decision, error context
- P1: Filter _extract_texts_from_messages to user-role messages only,
preventing system/assistant content from being injected into texts
- P1: Strengthen test_redact_updates_structured_messages assertion from
weak `in` check to strict equality, catching the injection bug
- P2: Use `result.get("decision") or "allow"` to handle explicit null
decision values (not just absent keys)
- P2: Wrap bare exception re-raise in GuardrailRaisedException so the
caller knows which guardrail failed (block_on_error=True path)
- P2: Add static Promptguard entry in guardrail_provider_map so the
preset works before populateGuardrailProviderMap is called
- Add test for explicit null decision treated as allow
* Fix black formatting: collapse f-string in error message
* fix(vertex_ai): support pluggable (executable) credential_source for WIF auth (#24700)
The WIF credential dispatch in load_auth() only handled identity_pool and
aws credential types. When credential_source.executable was present (used
for Azure Managed Identity via Workload Identity Federation), it fell
through to identity_pool.Credentials which rejected it with MalformedError.
Add dispatch to google.auth.pluggable.Credentials for executable-type
credential sources, following the same pattern as the existing identity_pool
and aws helpers.
Fixes authentication for Azure Container Apps → GCP Vertex AI via WIF
with executable credential sources.
* feat(logging): add component and logger fields to JSON logs for 3rd p… (#24447)
* feat(logging): add component and logger fields to JSON logs for 3rd party filtering
* Let user-supplied extra fields win over auto-generated component/logger, tighten test assertions
* Feat - Add organization into the metrics metadata for org_id & org_alias (#24440)
* Add org_id and org_alias label names to Prometheus metric definitions
* Add user_api_key_org_alias to StandardLoggingUserAPIKeyMetadata
* Populate user_api_key_org_alias in pre-call metadata
* Pass org_id and org_alias into per-request Prometheus metric labels
* Add test for org labels on per-request Prometheus metrics
* chore: resolve test mockdata
* Address review: populate org_alias from DB view, add feature flag, use .get() for org metadata
* Add org labels to failure path and verify flag behavior in test
* Fix test: build flag-off enum_values without org fields
* Gate org labels behind feature flag in get_labels() instead of static metric lists
* Scope org label injection to metrics that carry team context, remove orphaned budget label defs, add test teardown
* Use explicit metric allowlist for org label injection instead of team heuristic
* Fix duplicate org label guard, move _org_label_metrics to class constant
* Reset custom_prometheus_metadata_labels after duplicate label assertion
* fix: emit org labels by default, remove flag, fix missing org_alias in all metadata paths
* fix: emit org labels by default, no opt-in flag required
* fix: write org_alias to metadata unconditionally in proxy_server.py
* fix: 429s from batch creation being converted to 500 (#24703)
* add us gov models (#24660)
* add us gov models
* added max tokens
* Litellm dev 04 02 2026 p1 (#25052)
* fix: replace hardcoded url
* fix: Anthropic web search cost not tracked for Chat Completions
The ModelResponse branch in response_object_includes_web_search_call()
only checked url_citation annotations and prompt_tokens_details, missing
Anthropic's server_tool_use.web_search_requests field. This caused
_handle_web_search_cost() to never fire for Anthropic Claude models.
Also routes vertex_ai/claude-* models to the Anthropic cost calculator
instead of the Gemini one, since Claude on Vertex uses the same
server_tool_use billing structure as the direct Anthropic API.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>
* fix(anthropic): pass logging_obj to client.post for litellm_overhead_time_ms (#24071)
When LITELLM_DETAILED_TIMING=true, litellm_overhead_time_ms was null for
Anthropic because the handler did not pass logging_obj to client.post(),
so track_llm_api_timing could not set llm_api_duration_ms. Pass
logging_obj=logging_obj at all four post() call sites (make_call,
make_sync_call, acompletion, completion). Add test to ensure make_call
passes logging_obj to client.post.
Made-with: Cursor
* sap - add additional parameters for grounding
- additional parameter for grounding added for the sap provider
* sap - fix models
* (sap) add filtering, masking, translation SAP GEN AI Hub modules
* (sap) add tests and docs for new SAP modules
* (sap) add support of multiple modules config
* (sap) code refactoring
* (sap) rename file
* test(): add safeguard tests
* (sap) update tests
* (sap) update docs, solve merge conflict in transformation.py
* (sap) linter fix
* (sap) Align embedding request transformation with current API
* (sap) fix after bot review
* (sap) fix after bot review
* (sap) fix after bot review
* (sap) fix after bot review
* (sap) fix after bot review
* (sap) fix after bot review
* (sap) fix after bot review
* (sap) fix after bot review
* (sap) fix after bot review
* (sap) fix after bot review
* (sap) fix after bot review
* (sap) fix after bot review
* (sap) mock commit
* (sap) run black formater
* (sap) add literals to models, add negative tests, fix test for tool transformation
* (sap) fix formating
* (sap) fix models
* (sap) fix after bot review
* (sap) fix after bot review
* (sap) fix after bot review
* (sap) fix after bot review
* (sap) fix after bot review
* (sap) fix after bot review
* (sap) commit for rerun bot review
* (sap) minor improve
* (sap) fix after bot review
* (sap) lint fix
* docs(sap): update documentation
* fix(sap): change creds priority
* fix(sap): change creds priority
* fix(sap): fix sap creds unit test
* fix(sap): linter fix
* fix(sap): linter fix
* linter fix
* (sap) update logic of fetching creds, add additional tests
* (sap) clean up code
* (sap) fix after review
* (sap) fix after bot review
* (sap) fix after bot review
* (sap) fix after bot review
* (sap) fix after bot review
* (sap) fix after bot review
* (sap) fix after bot review
* (sap) fix after bot review
* (sap) fix after bot review
* (sap) fix after bot review
* (sap) fix after bot review
* (sap) fix after bot review
* (sap) add a possibility to put the service key by both variants
* (sap) fix after bot review
* (sap) fix after bot review
* (sap) fix after bot review
* (sap) update test
* (sap) update service key resolve function
* (sap) run black formater
* (sap) fix validate credentials, add negative tests for credential fetching
* (sap) fix validate credentials, add negative tests for credential fetching
* (sap) fix after bot review
* (sap) fix after bot review
* (sap) fix after bot review
* (sap) fix after bot review
* (sap) lint fix
* (sap) lint fix
* feat: support service_tier in gemini
* chore: add a service_tier field mapping from openai to gemini
* fix: use x-gemini-service-tier header in response
* docs: add service_tier to gemini docs
* chore: add defaut/standard mapping, and some tests
* chore: tidying up some case insensitivity
* chore: remove unnecessary guard
* fix: remove redundant test file
* fix: handle 'auto' case-insensitively
* fix: return service_tier on final steamed chunk
* chore: black
* feat: enable supports_service_tier to gemini models
* Fix get_standard_logging_metadata tests
* Fix test_get_model_info_bedrock_models
* Fix test_get_model_info_bedrock_models
* Fix remaining tests
* Fix mypy issues
* Fix tests
* Fix merge conflicts
* Fix code qa
* Fix code qa
* Fix code qa
* Fix greptile review
---------
Co-authored-by: michelligabriele <gabriele.michelli@icloud.com>
Co-authored-by: Josh <36064836+J-Byron@users.noreply.github.com>
Co-authored-by: mubashir1osmani <mubashir.osmani777@gmail.com>
Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>
Co-authored-by: milan-berri <milan@berri.ai>
Co-authored-by: Alperen Kömürcü <alperen.koemuercue@sap.com>
Co-authored-by: Vasilisa Parshikova <vasilisa.parshikova@sap.com>
Co-authored-by: Lin Xu <lin.xu03@sap.com>
Co-authored-by: Mark McDonald <macd@google.com>
Co-authored-by: Sameer Kankute <sameer@berri.ai>
Allow JWT tokens matching routing_overrides to use OAuth2 introspection without enabling global OAuth2 while keeping OAuth2 routing limited to LLM/info routes. Add regression coverage for management-route boundary and tighten opaque-token assertions; update docs to reflect selective-mode route scope.
Made-with: Cursor
Pin all cosign public key references to the immutable commit hash
(0112e53) that first introduced the key, instead of fetching it from
the release tag. This addresses the concern that an attacker with push
access could replace the key on main/tags and re-sign tampered images.
Docs now show two verification methods: commit hash (recommended) and
release tag (convenience), with explanation of why the hash is stronger.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
