Jj SFdZddlmZddlZddlZddlZddlmZddlm Z m Z ddl m Z ddl mZddlmZmZmZmZmZmZmZmZmZmZmZmZmZdd lmZmZdd l mZ!dd l"m#Z#ejHe%Z&e Gd d Z'Gdde(Z)Gdde)Z*GddZ+y)uRModel Gateway — unified interface for the Model Layer. The Model Gateway provides a single entry point for: - Chat/completion requests with automatic routing and fallback - Embedding generation - Model and provider registry management - Model intelligence (discovery, validation, classification) - Health checking across all providers ) annotationsN)Iterator) dataclassfield)Any)ModelEgressGate) LiteLLMClient LiteLLMConfigModelIntelligence ModelRecord ModelRegistry ModelRouter OllamaClient OllamaConfigProviderRecordProviderRegistryProviderStatus ProviderTrust RouterConfig)LlamaCppClientLlamaCppConfig)r) SecurityErrorceZdZUdZdZded<edZded<d Zd ed <d Z d ed <d Z d ed <dZ ded<dZ ded<dZ ded<dZded<dZded<y)ModelGatewayConfigz Configuration for Model Gateway.litellmstrdefault_providerc ddgS)Nollama llama.cppr!W/home/oem/CyberBrain26-unified/src/cyberbrain26/control_plane/model_gateway/__init__.pyzModelGatewayConfig.0s 8[BYr")default_factory list[str]fallback_providersTbool prefer_local prefer_free free_only local_firststrategyNzLiteLLMConfig | NonezOllamaConfig | NonerzLlamaCppConfig | None llama_cppzRouterConfig | Nonerouter)__name__ __module__ __qualname____doc__r__annotations__rr'r)r*r+r-rrr.r/r!r"r#rr+su*%c%$):Y$Z ZL$KIt!Hc!$(G !("&F &'+I$+"&F &r"rceZdZdZy)ModelGatewayErrorz(Base exception for Model Gateway errors.Nr0r1r2r3r!r"r#r6r6<s2r"r6ceZdZdZy)ModelNotAvailableErrorz+Raised when no suitable model is available.Nr7r!r"r#r9r9@s5r"r9ceZdZdZ d&ddd d'dZd(dZd(dZddddd dd  d)d Zd*d Z d+d Z ddd d,dZ d-dZ d& d.dZ dd d/dZ d0dZd1dZd2dZd3dZd4dZd5dZd6dZd3dZd&d7dZd8dZd9dZddddddd d:d Zdddd! d;d"Zdd%Zy)? ModelGatewayz'Unified Model Gateway for CyberBrain26.Nz model-gateway) egress_gateagent_idc|xs t|_t|jj|_t |jj |_t|jj|_ t|_ t|_t|j|j|_t#|j|j|jj$xs5t'|jj(|jj*|j |_|j-||_t1||_y)N)model_registryprovider_registry)r)r*)r?r@configmodel_intelligence)r<)rrAr rlitellm_clientrr ollama_clientrr.llama_cpp_clientr r?rr@r rBrr/rr)r*_register_builtin_providersr=r model_egress)selfrAr<r=s r#__init__zModelGateway.__init__Gs4 2 4 +DKK,?,?@)$++*<*<= .t{{/D/D E-:_3C3E"3.."44#  ".."44;;%%![[55 KK33* $66   ((* ! + Dr"ctdtjddddddgdtjddddd}|j j |tdtjddddddgd tjddddd }|j j |td tjdddd ddd gtjd dddd }|j j ||jy) NrTFzhttps://api.litellm.ai) z gemini-flashz gemini-prozgpt-4oz gpt-4o-minizclaude-3-5-sonnetzclaude-3-haikuz mistral-largezmistral-mediumz groq-llama-3z groq-mixtral)providerstatus free_tierregistration_required card_requiredapi_baseapi_key_required api_availableavailable_modelstrustsourcefreeno_cardapi_availabilityofficial_sourcerzhttp://127.0.0.1:11434)llama3zllama3.1mistralmixtralphi3gemma2zqwen2.5)rKrLrMrNrOrPrQrRrSrTrUrVrWlocal_availabilityrXr zhttp://127.0.0.1:8080) rrACTIVErOFFICIALr@register COMMUNITY_register_builtin_models)rHlitellm_providerollama_providerllama_cpp_providers r#rFz(ModelGateway._register_builtin_providersns')!(("'-!  ((! 5 8 ''(89(!(("'-" ))#!/ 2 ''8+ !(("',")]))#! " ''(:; %%'r"cddlm}m}m}m}m}m}|jjdxs tdgjD]]}|jj||d|j|jdd|j|j |ddd _|jjd xs td gjD]\}|jj||d |j|jdd|j|j"|dd  ^|jjd xs td gjD]Y}|jj||d |j|jdd|j|j"| [y) Nr)ModelCapabilitiesModelFreeStatusr ModelStatus ModelTrust ModelTyper)rKrSTF) tool_calling reasoningcoding) model_idrK model_typerLrRlocal_available free_statusrT capabilitiesr)rnrpr )cyberbrain26.model_layerrirjr rkrlrmr@getrrSr?rbCHATr`FREErarc)rHrirjr rkrlrmrqs r#rdz%ModelGateway._register_builtin_modelss  //33I>y.ZcvxByLL H    ( (%&(~~&--"&$) / 4 4$--!2%)"&#"  &//33H=wYatvAwJJ H    ( (%%(~~&--"&$( / 4 4$..!2%)#"   $//33K@}N\gz|D}PP H    ( (%((~~&--"&$( / 4 4$..!2!4   r"FT) temperature max_tokensstreamr)fallbackclassificationsc @|j||} |j| j||| |j| |||||S#t$rK} |r2t j d| |j||||||cYd} ~ Std|d| | d} ~ wwxYw)Nrqr)z)Primary model failed, trying fallback: %s)r~zChat failed for model z: ) _resolve_route_authorize_egressrK _execute_chat Exceptionloggerwarning_fallback_chatr9) rHmodelmessagesrzr{r|r)r}r~routeexcs r#chatzModelGateway.chats##U#N u~~uhP %%uh Z  JCP**8[*f$3+)(r#7  s#A B.BBBBcn|jj|}t|duxr |jS)zWhether the registry records ``model_id`` as free-eligible. Unknown models are treated as not free: the free-only guarantee must fail closed when the registry has no record to consult. N)r?rwr(is_free)rHrqrecords r#_model_is_freezModelGateway._model_is_frees3 $$((2F$&96>>::r"c 0|jjy|s tdtj|dd}|jj |j ||tj|jdj|y)aClear the request with the egress gate before any provider call. Fail-closed: a missing classification is refused rather than assumed public. No gate configured means no check, which is the pre-F4 behaviour and is explicit at construction time. NzKmodel egress requires an explicit classification; refusing to assume PUBLICFT) ensure_ascii sort_keyszutf-8)r= provider_idrqpayload_sha256r~) rG _egress_gaterjsondumpsauthorize_model_callr=hashlibsha256encode hexdigest)rH provider_namerrr~payloads r#rzModelGateway._authorize_egress#s    ) ) 1    **XETJ ..]]%">>'..*ABLLN+ / r"rc.|rR|jj|}|r5|jr)|jj|j}|r|S|s|j j rB|jj}|D]#}|jtjk(s!|cS|jjd}|r|jtjk(r|S|jj}|r|dStd)NrrzNo active providers available) r?rwrsr@rKrAr) list_localrL_ProviderStatusr` list_activer9) rHrqr)rrKlocal_providerspractives r#rzModelGateway._resolve_routeBs ''++H5E..1155ennE#O 4;;33"44??AO$ 88555H ((,,Y7 w~~)?)??N''335 !9 $%DEEr"cL|jdk(r |jj|||||S|jdk(r |jj|||||S|jdk(r |jj|||||St d|j)Nr)rrrzr{r|r)rrrz num_predictr|r Unknown provider: )rKrCrrDrEr9)rHrKrrrzr{r|s r#rzModelGateway._execute_chat_s    )&&++!'% ,   ( *%%**!'& +   + -((--!'% . )+=h>O>O=P)QR Rr"c d|jj}|D]}|jj|} | s!| jt j k(s?|jjr)|j|stjd||~ |j| j||| |j| |||||cSt!d|z#t$rt$rtjd|YwxYw#t$rtjd|YwxYw)Nz;Fallback provider %s refused: model %r is not free-eligiblez+Fallback provider %s blocked by egress gatezFallback provider %s failedz,All fallback providers exhausted for model: )rAr'r@rwrLrr`r+rrrrrKrrrr9) rHrrrzr{r|r~r'rrKs r#rzModelGateway._fallback_chats:"[[;;/ M--11-@HHOO/E/EE;;((1D1DU1KNNU%  ** ))5(O-- %; F3 D% :U B  #% NNE} !NN5} s$C7D (D D  D/.D/)r)c|s|jjrS|jj}|D]4}|jt j k(s!|j|||cS|jjd}|r0|jt j k(r|j|||Std)NrzNo provider for embeddings) rAr)r@rrLrr`_execute_embedrwr9)rHrtextsr)rrrs r#embedzModelGateway.embeds 4;;33"44??AO$ @88555..q%?? @((,,Y7 w~~)?)??&&wu= =$%ABBr"c:|jdk(r|jj||S|jdk(r|jj||S|jdk(r|jj||St d|j)Nr)rrrr r)rKrCrrDrEr9)rHrKrrs r#rzModelGateway._execute_embeds    )&&,,5,F F   ( *%%++%u+E E   + -((..U%.H H(+=h>O>O=P)QR Rr"c8|jj|SN)r?rbrHrs r#register_modelzModelGateway.register_models""++E22r"c8|jj|Sr)r?rw)rHrqs r# get_modelzModelGateway.get_models""&&x00r"c6|jjSr)r?list_allrHs r# list_modelszModelGateway.list_modelss""++--r"c8|jj|Sr)r@rb)rHrKs r#register_providerzModelGateway.register_providers%%..x88r"c8|jj|Sr)r@rwrHrs r# get_providerzModelGateway.get_providers%%))-88r"c6|jjSr)r@rrs r#list_providerszModelGateway.list_providerss%%..00r"c6|jjSr)rBdiscover_from_providersrs r#discover_modelszModelGateway.discover_modelss&&>>@@r"c8|jj|Sr)rB run_pipeline)rHmodelss r#run_validation_pipelinez$ModelGateway.run_validation_pipelines&&33F;;r"cL|jj|jSr)rBclassifyvaluers r#classify_modelzModelGateway.classify_models&&//6<<D//99;< %)KK$J$J$L #DKK$A$A$C D # ..779 H%%h&7&78F:@:M:M:O $X%6%67 % ( 1 1(; $X%6%67 r"cj|dk(r |jS|dk(r |jS|dk(r |jSy)Nrrr )rCrDrErs r#rzModelGateway._get_clientKsA I %&& & h &%% % k )(( (r"c|jjD]2}|j|j}|s!|j s2yy)NTF)r@rrrK is_available)rHrKrs r# is_healthyzModelGateway.is_healthyTsL..::< H%%h&7&78F&--/ r"r)rAzModelGatewayConfig | Noner<z Any | Noner=rreturnNone)rr)rrrlist[dict[str, str]]rz float | Noner{ int | Noner|r(r) bool | Noner}r(r~frozenset[Any] | Nonerstr | Iterator[dict[str, Any]])rqrrr() rrrrrrr~rrr)rq str | Noner)rrr)rKrrrrrrzrr{rr|r(rr)rrrrrzrr{rr|r(r~rrr)rrrr&r)rrlist[list[float]])rKrrrrr&rr)rr rr )rqrrzModelRecord | None)rzlist[ModelRecord])rKrrr)rrrzProviderRecord | None)rzlist[ProviderRecord])rzlist[ModelRecord] | Nonerlist[dict[str, Any]])rr rr)rr rfloat)rqrrudict[str, bool] | Noner)rr*rr-rr+rrdict[str, Any])rqrrurr+rrr)rr)rrrr)rr()r0r1r2r3rIrFrdrrrrrrrrrrrrrrrrrrrrrrrr!r"r#r;r;Ds1-1%E#'' %E)%E %E  %E  %ENM(^<F%)!%$(15' "   "/ (>;   '  /    D $$( FF" F  F:"S "S"S' "S " "S  "S"S ("SV26. . '. " .  .  . /.  (. j%) CCC " C  C& S  S S S  S31.991A<=: $/3$(#'#!%" " - " " " ! " " "  " N $/3!%   -      08r"r;),r3 __future__rrrloggingcollections.abcr dataclassesrrtypingr5cyberbrain26.control_plane.model_gateway.model_egressrrvr r r r r rrrrrrrr)cyberbrain26.model_layer.llama_cpp_clientrr*cyberbrain26.model_layer.provider_registryr&cyberbrain26.security_plane.permissionr getLoggerr0rrrr6r9r;r!r"r#rs# $(QUX@   8 $  ' '  ' 3 36.6UUr"