scriptoriumadmin/litellm
1
0
Fork 0
mirror of https://github.com/BerriAI/litellm.git synced 2026-10-10 03:28:53 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions 85

add granite guardian and input filtering target selector

Browse source
This commit is contained in:
Yamac Ay 2026-10-02 10:05:26 +02:00
parent 1726d4a26b
commit fd191dae23
No known key found for this signature in database
GPG key ID: D113B438819CA628
1 changed files with 164 additions and 19 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

183
litellm/llms/sap/chat/models.py
View file

@ -405,6 +405,43 @@ class MaskingModuleConfig(BaseModel):
return self
class AfterLastRoleTargetSelector(BaseModel):
"""
Scopes input filtering to all messages after the last message with a given role.
If no messages remain after applying this filter, filtering is skipped entirely
rather than raising an error.
Args:
after_last_role: The role used as the anchor. All messages that come after
the last occurrence of this role in the combined message list
(messages_history prepended to template) will be filtered.
One of: 'system', 'user', 'assistant', 'developer', 'tool'.
"""
after_last_role: Literal["system", "user", "assistant", "developer", "tool"]
class LastMessagesTargetSelector(BaseModel):
"""
Scopes input filtering to the last N messages of the combined message list.
Args:
last_messages: Number of messages from the end of the combined message list
(messages_history prepended to template) to include in filtering.
Must be >= 1 (0 is not allowed and returns 400 Bad Request).
If the value exceeds the total number of messages, all messages are filtered.
"""
last_messages: int = Field(ge=1)
# InputFilterTargetSelector is a discriminated union: exactly one of the two selector
# shapes must be present. The oneOf contract from the spec is enforced at instantiation
# time because each shape carries a field the other does not.
InputFilterTargetSelector = AfterLastRoleTargetSelector | LastMessagesTargetSelector
class AzureThreshold(int, Enum):
"""
Enumerates the threshold levels for the Azure Content Safety service.
@ -498,33 +535,25 @@ class LlamaGuard38bFilter(BaseModel):
Specific implementation of ContentFilter for Llama Guard 3. Llama Guard 3 is a
Llama-3.1-8B pretrained model, fine-tuned for content safety classification.
At least one category must be set to True (spec: minProperties: 1).
.. deprecated::
Will be removed 2027-09-20. Use :class:`GraniteGuardian41` instead.
Args:
violent_crimes: Responses that enable, encourage, or endorse the commission of violent crimes.
non_violent_crimes: Responses that enable, encourage, or endorse the commission of non-violent crimes.
sex_crimes: Responses that enable, encourage, or endorse the commission of sex-related crimes.
child_exploitation: Responses that contain, describe, enable or endorse the sexual abuse of children.
defamation: Responses that are both verifiably false and likely to injure a living person's reputation.
specialized_advice: Responses that contain specialized financial, medical or legal advice.
privacy: Responses that contain sensitive or nonpublic personal information.
intellectual_property: Responses that may violate the intellectual property rights of any third party.
indiscriminate_weapons: Responses that enable, encourage, or endorse the creation of indiscriminate weapons.
hate: Responses that demean or dehumanize people on the basis of their sensitive, personal characteristics.
self_harm: Responses that enable, encourage, or endorse acts of intentional self-harm.
sexual_content: Responses that contain erotica.
elections: Responses that contain factually incorrect information about electoral systems and processes.
code_interpreter_abuse: Responses that seek to abuse code interpreters.
"""
@ -543,15 +572,111 @@ class LlamaGuard38bFilter(BaseModel):
elections: bool = Field(default=False)
code_interpreter_abuse: bool = Field(default=False)
@model_validator(mode="after")
def enforce_min_one_category(self) -> "LlamaGuard38bFilter":
"""At least one category must be enabled (spec: minProperties: 1)."""
if not any(
[
self.violent_crimes,
self.non_violent_crimes,
self.sex_crimes,
self.child_exploitation,
self.defamation,
self.specialized_advice,
self.privacy,
self.intellectual_property,
self.indiscriminate_weapons,
self.hate,
self.self_harm,
self.sexual_content,
self.elections,
self.code_interpreter_abuse,
]
):
raise ValueError(
"LlamaGuard38bFilter requires at least one category set to True."
)
return self
class LlamaGuard38bFilterConfig(BaseModel):
type_: Literal["llama_guard_3_8b"] = Field(default="llama_guard_3_8b", alias="type")
config: LlamaGuard38bFilter
target_selector: InputFilterTargetSelector | None = None
class GraniteGuardian41Categories(BaseModel):
"""
Content categories evaluated by IBM Granite Guardian 4.1.
At least one category must be set to True (minProperties: 1 in the spec).
Granite Guardian issues a separate inference call per enabled category;
for most use cases enabling only ``harm`` is recommended as a catch-all.
Args:
harm: Catch-all criterion for generally harmful content.
social_bias: Detect prejudice or discrimination based on identity or
protected characteristics.
jailbreak: Detect attempts to manipulate the model into producing harmful
or otherwise undesired outputs.
violence: Detect content promoting or depicting physical, mental, or
sexual harm.
profanity: Detect offensive language or insults.
sexual_content: Detect explicit or suggestive material of a sexual nature.
unethical_behavior: Detect content describing actions that violate moral
or legal standards.
"""
harm: bool = Field(default=False)
social_bias: bool = Field(default=False)
jailbreak: bool = Field(default=False)
violence: bool = Field(default=False)
profanity: bool = Field(default=False)
sexual_content: bool = Field(default=False)
unethical_behavior: bool = Field(default=False)
@model_validator(mode="after")
def enforce_min_one_category(self) -> "GraniteGuardian41Categories":
"""At least one category must be enabled (spec: minProperties: 1)."""
if not any(
[
self.harm,
self.social_bias,
self.jailbreak,
self.violence,
self.profanity,
self.sexual_content,
self.unethical_behavior,
]
):
raise ValueError(
"GraniteGuardian41Categories requires at least one category set to True."
)
return self
class GraniteGuardian41(BaseModel):
"""
Configuration for IBM Granite Guardian 4.1 filter provider.
Args:
enable_reasoning: Enable reasoning (think) mode. When True, the model returns
an explanation alongside each verdict, e.g.
``{'verdict': True, 'reasoning': '...'}``. Applies to every configured
category. Defaults to False.
categories: Content criteria to evaluate. At least one category must be
enabled. Granite Guardian issues a separate inference call per category;
using only ``harm`` is recommended as a catch-all to minimise latency.
"""
enable_reasoning: bool = Field(default=False)
categories: GraniteGuardian41Categories
class AzureContentSafetyInputFilterConfig(BaseModel):
type_: Literal["azure_content_safety"] = Field(default="azure_content_safety", alias="type")
config: AzureContentSafetyInput | None = None
target_selector: InputFilterTargetSelector | None = None
class AzureContentSafetyOutputFilterConfig(BaseModel):
@ -559,6 +684,22 @@ class AzureContentSafetyOutputFilterConfig(BaseModel):
config: AzureContentSafetyOutput | None = None
class GraniteGuardianFilterConfig(BaseModel):
"""
Filter configuration for the IBM Granite Guardian 4.1 provider.
Args:
type_: Provider discriminator — always ``'granite_guardian_4_1'``.
config: Category and reasoning settings for Granite Guardian.
target_selector: Optional selector to scope filtering to a subset of the
combined message list. When absent, all input content is filtered.
"""
type_: Literal["granite_guardian_4_1"] = Field(default="granite_guardian_4_1", alias="type")
config: GraniteGuardian41
target_selector: InputFilterTargetSelector | None = None
class FilteringStreamOptions(BaseModel):
"""
overlap: Number of characters that should be additionally sent to content filtering services
@ -572,22 +713,26 @@ class InputFiltering(BaseModel):
"""Module for managing and applying input content filters.
Args:
filters: List of ContentFilter objects to be applied to input content.
filters: List of filter provider configurations to be applied to input content.
Supported providers: Azure Content Safety, Llama Guard 3 8B (deprecated),
and IBM Granite Guardian 4.1.
"""
filters: list[AzureContentSafetyInputFilterConfig | LlamaGuard38bFilterConfig] = Field(min_length=1)
filters: list[AzureContentSafetyInputFilterConfig | LlamaGuard38bFilterConfig | GraniteGuardianFilterConfig] = Field(min_length=1)
class OutputFiltering(BaseModel):
"""Module for managing and applying output content filters.
Args:
filters: List of ContentFilter objects to be applied to output content.
stream_options: Module-specific streaming options.
filters: List of filter provider configurations to be applied to output content.
Supported providers: Azure Content Safety, Llama Guard 3 8B (deprecated),
and IBM Granite Guardian 4.1.
stream_options: Module-specific streaming options. Ignored when streaming is
disabled.
"""
filters: list[AzureContentSafetyOutputFilterConfig | LlamaGuard38bFilterConfig] = Field(min_length=1)
filters: list[AzureContentSafetyOutputFilterConfig | LlamaGuard38bFilterConfig | GraniteGuardianFilterConfig] = Field(min_length=1)
stream_options: FilteringStreamOptions | None = None