tPj3dZddlmZmZmZmZddlmZddlm Z ddl m Z m Z ddl mZddlmZmZddlmZdd lmZmZmZd d lmZer dd lmZdd lmZGddee Zy)z5 OpenAI Moderation Guardrail Integration for LiteLLM ) TYPE_CHECKINGFinalLiteralOptional) HTTPException)verbose_proxy_logger)CustomGuardraillog_guardrail_information)Logging)get_async_httpx_clienthttpxSpecialProvider)GuardrailEventHooks)GenericGuardrailAPIInputsGuardrailStatusGuardrailTracingDetail)OpenAIGuardrailBaseOpenAIModerationResponse)GuardrailConfigModelceZdZdZ d&dededzdedzdeddzdedzd edzf fd Zd edzfd Z d ed dfdZ d'dZ e d(de dededdedd e f dZ d&dedzdededzdedzdedzdedzdedzfdZ d)dedededzdedzdedzdedzf d Zed!eezezd edzfd"Zed ed#dzfd$Zed eefd%ZxZS)*OpenAIModerationGuardraila LiteLLM Built-in Guardrail for OpenAI Content Moderation. This guardrail scans prompts and responses using the OpenAI Moderation API to detect harmful content, including violence, hate, harassment, self-harm, sexual content, etc. Configuration: guardrail_name: Name of the guardrail instance api_key: OpenAI API key api_base: OpenAI API endpoint model: OpenAI moderation model to use default_on: Whether to enable by default Nguardrail_nameapi_keyapi_basemodel)omni-moderation-latestztext-moderation-lateststreaming_end_of_stream_onlystreaming_sampling_ratec t|d |t|jd|t t j |_|xs|j|_ |xsd|_ |xsd|_ |dn||_ |dn||_ |js tdtj d ||jy) z/Initialize OpenAI Moderation guardrail handler.)rsupported_event_hooks) llm_providerzhttps://api.openai.com/v1rNFzlOpenAI Moderation: api_key is required. Set OPENAI_API_KEY environment variable or pass it in configuration.z:Initialized OpenAI Moderation Guardrail: %s with model: %s)super__init__listget_supported_event_hooksr r GuardrailCallback async_handler _get_api_keyrrrrr ValueErrorrdebug) selfrrrrrrkwargs __class__s /home/oem/CyberBrain26-unified/.venv/lib/python3.12/site-packages/litellm/proxy/guardrails/guardrail_hooks/openai/moderations.pyr&z"OpenAIModerationGuardrail.__init__1s  )"&t'E'E'G"H  4AUAgAgh5$"3"3"5  ?$? RWRs[s 29E?[ )2I1PAVm$||~  "" H.Z^ZdZd returncddl}ddl}ddlm}|jj dxs$|j xs|jxs|dS)z?Get API key from environment variables or litellm configurationrN)get_secret_strOPENAI_API_KEY)oslitellmlitellm.secret_managers.mainr5environgetr openai_key)r.r7r8r5s r1r+z&OpenAIModerationGuardrail._get_api_keyYsM? JJNN+ , 0 0!! 0./  r2 input_textrcK|j|d}tjd||jj |j dd|j dd|d{}tjd |j|jd k7r$t|jd |jd  ddl m }|di|jS7ww)z> Make a request to the OpenAI Moderation API. )rinputz#OpenAI Moderation guard request: %sz /moderationszBearer zapplication/json) Authorizationz Content-Type)urlheadersjsonNz$OpenAI Moderation guard response: %sz$OpenAI Moderation API request failed)errordetails status_codedetailrrr$) rrr-r*postrrrCrHrtextlitellm.types.llms.openair)r.r= request_bodyresponsers r1async_make_requestz,OpenAIModerationGuardrail.async_make_requestgs)- ZH ""#H,W $ 2 2 7 7==/.#*4<<.!9 2 !8!   ""#I8==?[   3 &$00C'}}  G':(--/::- sA$C &C'A8C cN|jsy|jd}|jr}g}|jr6|jjD]\}}|s |j |||j xsid}t jd|tdd|dy) zk Check if the moderation response indicates harmful content and raise exception if needed. Nr)violated_categoriescategory_scoresz5OpenAI Moderation: Content flagged for violations: %siz!Violated OpenAI moderation policy)rEmoderation_resultrG) resultsflagged categoriesitemsappendrRrwarningr)r.moderation_responseresultrQcategory is_violatedviolation_detailss r1_check_moderation_resultz2OpenAIModerationGuardrail._check_moderation_results#** +33A6 >>)+   -3->->-D-D-F=)Hk"+228<= (;#)#9#9#?R(  ! ( (G!   @): % r2inputs request_data input_type)requestrN logging_objLiteLLMLoggingObjcKd}|jdx}r|j|}|s$|jdx}rdj|}|stjd|S|j |d{}t |tr-|jdxsi} | |d<|j| d<|j||S7Tw) a Apply OpenAI moderation guardrail using the unified guardrail interface. This method is called by the UnifiedLLMGuardrails system for all endpoint types (chat completions, embeddings, responses API, etc.). Args: inputs: GenericGuardrailAPIInputs containing texts and/or structured_messages request_data: The original request data input_type: Whether this is a "request" (pre-call) or "response" (post-call) logging_obj: Optional logging object Returns: The inputs unchanged (moderation doesn't modify content, only blocks) Raises: HTTPException: If content violates moderation policy Nstructured_messagestexts z8OpenAI Moderation: No text content to moderate in inputs)r=metadata_openai_moderation_response) r;get_user_promptjoinrr-rO isinstancedict model_dumpr_) r.r`rarbrdtext_to_moderatergrhrZrjs r1apply_guardrailz)OpenAIModerationGuardrail.apply_guardrails6(,#)**-B"C C  C#334GH   7++u+#'99U#3  & &'a bM,0+B+BN^+B+_%_ lD )*..z:@bH'/L $6I6T6T6VH2 3 %%&9: &`sA;C=C>ACrN start_timeend_timeduration event_typeoriginal_inputsc t|tr|jdxsi}||d<ni}|jdd} |j | |d|||||j | |S)z Override to log the full OpenAI Moderation API response instead of the decorator's simplified "allow"/"mask" string. Follows the Model Armor pattern (model_armor.py:325-360). rjrkallowsuccessguardrail_json_responseraguardrail_statusrursrtrvtracing_detail)rnror;pop:add_standard_logging_guardrail_information_to_request_data_build_tracing_detail) r.rNrarsrtrurvrwrjguardrail_responses r1_process_responsez+OpenAIModerationGuardrail._process_responses lD )#'' 39rH'/L $H %-LL1NPW$X GG$6%&!!556HI H r2ec |j|rdnd}t|tr|jdxsi}||d<ni}|j d|} |j | |||||||j | |)z Override to log the full OpenAI Moderation API response on error instead of the stringified exception. guardrail_intervenedguardrail_failed_to_respondrjrkr{)_is_guardrail_interventionrnror;rrr) r.rrarsrtrurvr}rjrs r1_process_errorz(OpenAIModerationGuardrail._process_error s'+&E&Ea&H "Nk  lD )#'' 39rH'/L $H=ELLIfhirsV;:!1T9 &BXz 3_z r2