Add vector_size parameter to QdrantSemanticCache and expose it through
the Cache facade as qdrant_semantic_cache_vector_size. This allows users
to use embedding models with dimensions other than the default 1536,
enabling cheaper/stronger models like Stella (1024d), bge-en-icl (4096d),
voyage, cohere, etc.
The parameter defaults to QDRANT_VECTOR_SIZE (env var or 1536) for
backward compatibility. When creating new collections, the configured
vector_size is used instead of the hardcoded constant.
Closes#9377
POST /access_group/new and PUT /access_group/{name}/update now accept an
optional model_ids list that targets specific deployments by their unique
model_id, instead of tagging every deployment that shares a model_name.
When model_ids is provided it takes priority over model_names, giving
API callers the same single-deployment precision that the UI already has
via PATCH /model/{model_id}/update.
Backward compatible: model_names continues to work as before.
Closes#21544
Add user_email field to LiteLLM_OrganizationMembershipTable with a model_validator
that populates it from the nested user object, and update the /organization/info
Prisma query to select user_email from the related user record.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
- Prefix credential name tags with "Credential: " to distinguish them
from user-defined tags when litellm_credential_name is injected
- Remove stale "new feature" banners from Organization, Customer, and
A2A usage views
- Add closable info banner to Tag usage view noting that reusable
credentials are automatically tracked and appear as "Credential: <name>"
* feat: add competitor name blocker guardrail
* fix: fix batch test endpoint for compliance playground
* fix(airline.py): add list of all known airlines to airline competitor name detector
prevent competitor discussion on company chatbot
* feat: ui tweaks for prod
Fields like input_cost_per_token contain "token" as a key segment,
which incorrectly matched the sensitive patterns list and caused values
like 3.6e-06 to be displayed as "3.60*******e-06" in the model info UI.
Add a non_sensitive_overrides set (defaulting to {"cost"}) so that any
key containing "cost" as a segment is never masked, regardless of other
matching patterns.
* Add keyword-based topic blocker implementation
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
* Add embedding-based topic blocker using MiniLM
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
* Add topic blocker package init with exports
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
* Add synthetic engine eval set (34 cases)
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
* Add investment questions eval set (207 cases)
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
* Add engine eval synthetic policy config
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
* Add engine keyword blocker eval results
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
* Add investment keyword blocker eval results
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
* Add investment embedding blocker eval results
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
* Add investment embedding MiniLM eval results
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
* Add investment embedding MPNet eval results (historical)
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
* Add investment TF-IDF eval results (historical)
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
* Add unified eval runner with confusion matrix reporting
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
* Add benchmarks comparison table in markdown
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
* Clean up topic blocker: remove unused blockers, add phrase_patterns to content filter
- Remove embedding_blocker.py, api_embedding_blocker.py, nli_blocker.py,
tfidf_blocker.py, onnx_blocker.py (heavy deps not in Docker, inferior accuracy)
- Remove airline_off_topic_restriction policy template and its test
- Fix __init__.py to only export DeniedTopic and TopicBlocker (no eager import crash)
- Add phrase_patterns support to ContentFilterGuardrail for regex-based paraphrase detection
- Rewrite denied_financial_advice.yaml with conditional matching (identifier + block word),
always-block keywords, phrase patterns, and exception phrases
- Clean up test_eval.py: only keyword blocker + content filter tests remain (no network calls)
- All 207 eval cases pass at 100% F1, 0 FP, 0 FN, <0.1ms latency
Addresses all Greptile review comments:
- Eager import crash (embedding deps) → fixed
- Undeclared dependencies → fixed (files deleted)
- lru_cache memory leak → fixed (file deleted)
- Real network calls in tests → fixed (embedding tests removed)
- Unused Dict import → already fixed
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
* Add LLM-as-judge eval and update BENCHMARKS.md
- Add TestInvestmentLlmJudgeGpt4oMini and TestInvestmentLlmJudgeClaude
test classes that use litellm.completion() to classify messages
- System prompt instructs LLM to act as airline chatbot content moderator
- Tests skip gracefully when API keys aren't set
- Update BENCHMARKS.md with production results table, historical comparison,
and instructions for running LLM judge evals
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
* Move evals and benchmarks to guardrail_benchmarks folder
Move eval runner, eval data (JSONL), and results from
tests/test_litellm/.../topic_blocker/ into the guardrail implementation
folder at litellm/.../litellm_content_filter/guardrail_benchmarks/.
This keeps benchmarks co-located with the guardrail code they test.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
* Remove standalone topic_blocker package, consolidate into content_filter
The standalone keyword_blocker.py was redundant with content_filter.py +
denied_financial_advice.yaml. Removed the entire topic_blocker/ package,
engine eval files, and old keyword blocker results. Simplified test_eval.py
to only test ContentFilter + LLM judge baselines.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
* Fix compliance playground batch scoring bug, add display_name support
The compliance playground was sending all texts in a single batch API call,
but the content filter raises HTTPException on the first blocked text. This
caused a single blocked/allowed result to be applied to all rows, producing
incorrect scores (e.g. 41% instead of 100%). Fix by sending each text
individually to get per-text results with progressive UI updates.
Also add display_name field support for category YAML files so
denied_financial_advice shows as "Denied Financial / Investment Advice"
in the UI dropdown.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
* Add block_investment CSV eval set and update benchmark result JSON
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
* address greptile review feedback (greploop iteration 1)
Fix stale test path in denied_financial_advice.yaml comment.
Other comments were on files already deleted in prior commits.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>
- Add new SpendLogsMetadata keys to ignored_keys in spend logs tests
(regression from ccecc10c82 which intentionally includes all keys)
- Mock PrismaManager.setup_database and should_update_prisma_schema in
proxy CLI tests to prevent real DB migrations from running in CI
- Use CliRunner(mix_stderr=False) to fix Click stream lifecycle issues
- Use unique UUID suffix for Redis TTL test keys to avoid stale state
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
The previous approach tried to inject litellm_credential_name as a tag
in get_standard_logging_object_payload, but the credential name was never
available in litellm_params because the Logging object is created by the
proxy BEFORE the Router selects a deployment. The credential name only
exists in the deployment's litellm_params, which is resolved later.
This fix injects the credential name as a tag in
Router._update_kwargs_with_deployment(), right alongside the existing
deployment-level tags mechanism. This ensures the credential name flows
through the normal metadata.tags pipeline.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
- RC1: Override _remove_key() in LLMClientCache to schedule aclose() on
evicted async clients instead of relying on GC
- RC2: Use passed connection_pool for URL configs instead of creating an
orphaned pool via from_url()
- RC3: Pass max_connections through to BlockingConnectionPool.from_url()
for URL configs, with input validation for invalid values
- RC5: Close sync redis_client in disconnect() with try/except guard
When a model has a litellm_credential_name, append it to request_tags during
logging so it flows into DailyTagSpend and becomes filterable in the Usage page.
- Add litellm_credential_name to _OPTIONAL_KWARGS_KEYS so it survives into
litellm_params during get_litellm_params() filtering
- Read credential name from litellm_params in get_standard_logging_object_payload()
and append to request_tags if not already present
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
* fix(responses): fix O(n²) CPU overhead in reasoning streaming path
stream_chunk_builder was called on every reasoning chunk, rebuilding the
entire response from all collected chunks each time. Replace with
incremental accumulation of reasoning_content parts, only joining at
reasoning end.
* fix(responses): eliminate per-chunk thread spawning in async streaming path
_process_chunk() called run_async_function() on every SSE chunk, which
when invoked from an async context spawns a thread + event loop per call.
Move the hook call out of _process_chunk into the callers: async __anext__
directly awaits it, sync __next__ uses run_async_function.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
* perf: reduce responses streaming CPU for text-only streams
* fix(test): replace deprecated claude-3-7-sonnet-latest in responses API test
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
* fix(test): replace deprecated claude-3-7-sonnet-latest in tool result fix test
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
* fix(test): replace deprecated claude-3-7-sonnet-latest in tool result empty call_id test
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>
Add attempted_retries and max_retries fields to SpendLogsMetadata so the
Logs page can display how many retries occurred for each request. The
router now injects retry tracking metadata before each make_call, which
flows through the logging pipeline into the spend logs metadata JSON.
The UI shows "Not Retried" when the first attempt succeeded, and
"N / M" (attempted / max) when retries occurred. The field is hidden
for requests that did not go through the router.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
ModelResponse.choices was typed as List[Union[Choices, StreamingChoices]] which
caused Pydantic serialization warnings and false linting errors. Now that
ModelResponseStream exists for streaming, narrow ModelResponse.choices to
List[Choices] and migrate all ModelResponse(stream=True) call sites to use
ModelResponseStream() instead.
Bedrock multimodal embedding models (Titan and Nova) were being costed
using the per-token text rate instead of the correct flat per-image rate
($0.00006/image). The pricing data was correct but never applied because
image_count was never populated in prompt_tokens_details.
Pass batch_data to Titan/Nova response transformers so they can count
image inputs and set PromptTokensDetailsWrapper(image_count=N) on Usage,
mirroring the existing Vertex AI pattern from PR #9623. Also fix the
text_tokens fallback in generic_cost_per_token to not override
text_tokens=0 when image_count > 0 (image-only requests).
* fix(model-pricing): add missing fireworks_ai model pricing for glm-4p7, minimax-m2p1, kimi-k2p5
Fireworks AI models called via short-form (fireworks_ai/<model>) were
reporting $0.00 cost because the pricing JSON lacked short-form entries.
The lookup fell through to the fireworks-ai-default bucket which has
zero cost.
Added 5 new entries to model_prices_and_context_window.json:
- fireworks_ai/accounts/fireworks/models/glm-4p7 (new long-form)
- fireworks_ai/accounts/fireworks/models/minimax-m2p1 (new long-form)
- fireworks_ai/glm-4p7 (new short-form)
- fireworks_ai/minimax-m2p1 (new short-form)
- fireworks_ai/kimi-k2p5 (new short-form; long-form already existed)
Pricing sourced from fireworks.ai model pages and pricing page.
* add cache_read_input_token_cost to kimi-k2p5 long-form entry for consistency
For gemini-3-pro-preview (not gemini-3-flash or gemini-3.1-pro-preview),
reasoning_effort="medium" maps to thinkingLevel="high" because the "medium"
thinking level is not available on that model variant.
Both test_reasoning_effort_maps_to_thinking_level_gemini_3 and
test_reasoning_effort_dict_format_gemini_3 had the correct comment
("medium -> high") but the wrong assertion (== "medium"). Fixed to
match the production code and the comments.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
The production error message was expanded when Sonnet 4.6 was also added as
a supported model for effort='max'. The test's match regex still referenced
the old "Claude Opus 4.6"-only message; update it to match the new
"Claude 4.6 models" wording.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Reset logo_path to default_logo when custom UI_LOGO_PATH file doesn't
exist, so the else branch at the bottom of get_image serves the default
logo instead of the non-existent custom path.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Add os.path.exists check before serving custom local logo so that a
non-existent UI_LOGO_PATH gracefully falls through to the cache/default
instead of causing a FileResponse error.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
The /get_image endpoint checked for cached_logo.jpg before reading the
UI_LOGO_PATH env var, so a pre-existing cache (e.g. baked into the base
Docker image) would always be served, ignoring the user's custom logo.
Move the UI_LOGO_PATH read before the cache check and serve local file
paths directly, bypassing the cache. The cache optimization is preserved
for HTTP URLs and the default logo where it is actually needed.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Empty system messages were skipped for Anthropic's system param but
not removed from the messages list, causing BadRequestError when
anthropic_messages_pt encountered the unsupported "system" role.
Fixes#21622
Pydantic v2's Union serializer for `List[Union[Choices, StreamingChoices]]`
tries both branches when serializing, emitting spurious
`PydanticSerializationUnexpectedValue` warnings (field count mismatch on
`Message` and type mismatch `Expected StreamingChoices but got Choices`).
Add a `WrapSerializer` on the `choices` field that serializes each item
individually via its own `model_dump()`, bypassing the Union dispatch
entirely while correctly propagating `exclude_none`, `exclude_unset`,
and `exclude_defaults` from the parent serialization context.
Regular team members could not see service accounts belonging to their
team. Additionally, former team members could still see service accounts
they created via the created_by filter after leaving the team.
- Add get_member_team_ids() to retrieve all teams a user belongs to
- Regular members now see team service accounts (user_id=NULL) but not
other members' personal keys
- Scope created_by filter to the user's current team memberships
- Add 7 unit tests covering the new visibility rules
Replace find_many + Python-side aggregation with a single SQL GROUP BY
query via query_raw in get_daily_activity_aggregated. This collapses
rows across entities (users/teams/orgs) in the database, reducing ~150k
rows to ~2-3k grouped rows before transfer to Python.
Also adds composite indexes (entity_id, date) to all 6 daily spend
tables for faster filtered queries.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
