tock-genai-core 1.2.0

No project description provided

tock-genai-core

Composants principaux d'IA générative : models, factories, gestion des erreurs utilisés dans les composants Gen AI python de tock.

Architecture

Le projet est structuré en trois composants principaux :

• Models : Contient les définitions des classes et modèles de données utilisés dans l'application
• Services/Factories : Regroupe la logique métier et les fonctions utilisées par les routes

Technologies

• Backend : Python
• Base de données :
  • pgvector pour la base de données vectorielle
• LLM & RAG :
  • Langchain pour l'orchestration
  • Support de guardrails
  • Reranker pour l'amélioration des résultats
  • Langfuse pour le monitoring

Providers disponibles

• LLMProvider (LLM)

  • TGI = "HuggingFaceTextGenInference"
  • OpenAI = "OpenAI"
  • Vllm = "Vllm"

• GuardrailProvider (Guardrail)

  • BloomZ = "BloomzGuardrail"

• EMProvider (Embedding)

  • BloomZ = "BloomzEmbeddings"
  • OpenAI = "OpenAI"
  • Vllm = "Vllm"

• VectorDBProvider (Database)

  • OpenSearch = "OPENSEARCH"
  • PGVector = "PGVECTOR"

• ContextualCompressorProvider (Contetual Compressor)

  • BloomZ = "BloomzRerank"

Settings

• Embedding

  • Classe parente

    BaseEMSetting:
        provider: EMProvider
        model: Optional[str]
        api_key: Optional[SecretKey]
        api_base: str
        pooling: Optional[str]
        space_type: Optional[str]
    

  • Classes enfants

    BloomZEMSetting(BaseEMSetting):
        provider: Literal[EMProvider.BloomZ]
    

    VLLMEMSetting(BaseEMSetting):
        provider: Literal[EMProvider.Vllm]
        model: str
    

    OpenAIEMSetting(BaseEMSetting):
        provider: Literal[EMProvider.OpenAI]
        api_base: str
        api_version: str
        deployment: str
    

• Contextual compressor

  • Classe parente

    BaseCompressorSetting:
        provider: ContextualCompressorProvider
        endpoint: str
        api_key: Optional[SecretKey]
    

  • Classe enfant

    BloomZCompressorSetting(BaseCompressorSetting):
        provider: Literal[ContextualCompressorProvider.BloomZ]
        min_score: float
        max_documents: Optional[int]
        label: Optional[str]
    

• Database

  • Classe parente

    BaseVectorDBSetting:
        index: Optional[str]
        provider: VectorDBProvider
        db_url: str
    

  • Classes enfants

    OpenSearchSetting(BaseVectorDBSetting):
        provider: Literal[VectorDBProvider.OpenSearch]
        username: SecretKey
        password: SecretKey
        use_ssl: bool
        verify_certs: bool
    

    class PGVectorSetting(BaseVectorDBSetting):
        provider: Literal[VectorDBProvider.PGVector]
        username: SecretKey
        password: SecretKey 
        db_name: str
        sslmode: Optional[str]
        namespace: str
    

• Guardrail

  • Classe parente

    BaseGuardrailSetting:
        provider: GuardrailProvider
        api_base: str
        max_score: Optional[float]
        api_key: Optional[SecretKey]
    

  • Classe enfant

    BloomZGuardrailSetting(BaseGuardrailSetting):
        provider: Literal[GuardrailProvider.BloomZ]
    

• Langfuse

  LangfuseSetting:
      host: Optional[str]
      public_key: Optional[SecretKey]
      secret_key: Optional[SecretKey]
      app_name: Optional[str]
      user_id: Optional[str]
      session_id: Optional[str]
  

• LLM

  • Classe parente

    BaseLLMSetting:
        provider: LLMProvider
        model: Optional[str]
        api_key: Optional[SecretKey]
        temperature: float
    

  • Classes enfants

    OpenAILLMSetting(BaseLLMSetting):
        provider: Literal[LLMProvider.OpenAI]
        api_base: str
        api_version: str
        deployment: str
    

    HuggingFaceTextGenInferenceLLMSetting(BaseLLMSetting):
        provider: Literal[LLMProvider.TGI]
        repetition_penalty: float
        max_new_tokens: int
        api_base: str
        streaming: bool
    

    VllmSetting(BaseLLMSetting):
        provider: Literal[LLMProvider.Vllm]
        api_base: str
        max_new_tokens: int
        additional_model_kwargs: Optional[Dict[str, Any]]
    

Fonctionnement

Chaque outil utilisé (database, embedding, llm, langfuse, ...) a besoin d'un certains nombre de paramètres qui sont référencés dans les models (classes de settings)

Ces classes sont ensuite héritées par des services ou des factories afin de pouvoir répondre au besoin.

Exemple de get_vector_db_factory qui crée une factory de base vectorielle basée sur le nom de l'application et les paramètres d'embedding fournis

from tock-genai-core import get_vector_db_factory
from tock-genai-core import PGVectorSetting, VLLMEMSetting
from tock-genai-core import DBSetting, EMSetting


db_settings = PGVectorSetting(
    index = "first_index",
    provider = "PGVECTOR",
    db_url = "127.0.0.1:XXXX",
    db_name = "rag_sandbox_db",
    sslmode = "disable",
    username = {
      type = "Raw",
      value = "admin"
    },
    password = {
      type = "Raw",
      value = "example"
    },
    namespace = "test-name"
)

em_settings = VLLMEMSetting(
    provider = "Vllm",
    model = "model_name",
    api_base = "https://continue.com/v1"
)



def function_name(db_settings: DBSetting, em_settings: EMSetting):

    # do somethings

    vector = get_vector_db_factory(db_settings: DBSetting, em_settings: BaseEMSetting)

    # do somethings