beekeeper-core 1.0.0__py3-none-any.whl

beekeeper/core/document/__init__.py

@@ -0,0 +1,3 @@
+from beekeeper.core.document.base import Document, DocumentWithScore
+
+__all__ = ["Document", "DocumentWithScore"]

beekeeper/core/document/base.py

@@ -0,0 +1,103 @@
+import uuid
+from abc import ABC, abstractmethod
+from hashlib import sha256
+from typing import Any, Dict, List, Optional, Union
+
+import numpy as np
+from pydantic.v1 import BaseModel, Field, validator
+
+
+class BaseDocument(ABC, BaseModel):
+    """Generic abstract interface for retrievable documents."""
+
+    id_: str = Field(
+        default_factory=lambda: str(uuid.uuid4()),
+        description="Unique ID of the document.",
+    )
+    metadata: Dict[str, Any] = Field(
+        default_factory=dict,
+        description="A flat dictionary of metadata fields.",
+    )
+    embedding: Optional[Union[List[float], np.ndarray]] = Field(
+        default_factory=None,
+        description="Embedding of the document.",
+    )
+
+    class Config:
+        arbitrary_types_allowed = True
+
+    @validator("metadata", pre=True)
+    def _validate_metadata(cls, v) -> Dict:
+        if v is None:
+            return {}
+        return v
+
+    @abstractmethod
+    def get_content(self) -> str:
+        """Get document content."""
+
+    def get_metadata(self) -> dict:
+        """Get metadata."""
+        return self.metadata
+
+    def get_embedding(self) -> List[float]:
+        """Get metadata."""
+        return self.embedding
+
+    @property
+    @abstractmethod
+    def hash(self) -> str:
+        """Get hash."""
+
+
+class Document(BaseDocument):
+    """Generic interface for data document."""
+
+    text: str = Field(default="", description="Text content of the document.")
+
+    @classmethod
+    def class_name(cls) -> str:
+        return "Document"
+
+    def get_content(self) -> str:
+        """Get the text content."""
+        return self.text
+
+    @property
+    def hash(self) -> str:
+        """Get document hash."""
+        return str(sha256(str(self.text).encode("utf-8", "surrogatepass")).hexdigest())
+
+
+class DocumentWithScore(BaseModel):
+    document: BaseDocument
+    score: Optional[float] = None
+
+    @classmethod
+    def class_name(cls) -> str:
+        return "DocumentWithScore"
+
+    def get_score(self) -> float:
+        """Get score."""
+        if self.score is None:
+            return 0.0
+        else:
+            return self.score
+
+    # #### pass through methods to BaseDocument ####
+    @property
+    def id_(self) -> str:
+        return self.document.id_
+
+    @property
+    def text(self) -> str:
+        if isinstance(self.document, Document):
+            return self.document.text
+        else:
+            raise ValueError("Must be a Document to get text.")
+
+    def get_content(self) -> str:
+        return self.document.get_content()
+
+    def get_metadata(self) -> str:
+        return self.document.get_metadata()

beekeeper/core/embeddings/__init__.py

@@ -0,0 +1,7 @@
+from beekeeper.core.embeddings.base import BaseEmbedding, Embedding, SimilarityMode
+
+__all__ = [
+    "BaseEmbedding",
+    "Embedding",
+    "SimilarityMode",
+]

beekeeper/core/embeddings/base.py

@@ -0,0 +1,66 @@
+from abc import ABC, abstractmethod
+from enum import Enum
+from typing import List
+
+import numpy as np
+from beekeeper.core.document import Document
+from beekeeper.core.schema import TransformerComponent
+from beekeeper.core.utils.pairwise import cosine_similarity
+
+Embedding = List[float]
+
+
+class SimilarityMode(str, Enum):
+    """Modes for similarity."""
+
+    COSINE = "cosine"
+    DOT_PRODUCT = "dot_product"
+    EUCLIDEAN = "euclidean"
+
+
+def similarity(
+    embedding1: Embedding,
+    embedding2: Embedding,
+    mode: SimilarityMode = SimilarityMode.COSINE,
+):
+    """Get embedding similarity."""
+    if mode == SimilarityMode.EUCLIDEAN:
+        return -float(np.linalg.norm(np.array(embedding1) - np.array(embedding2)))
+
+    elif mode == SimilarityMode.DOT_PRODUCT:
+        return np.dot(embedding1, embedding2)
+
+    else:
+        return cosine_similarity(embedding1, embedding2)
+
+
+class BaseEmbedding(TransformerComponent, ABC):
+    """An interface for embedding models."""
+
+    @classmethod
+    def class_name(cls) -> str:
+        return "BaseEmbedding"
+
+    @abstractmethod
+    def get_text_embedding(self, query: str) -> Embedding:
+        """Get query embedding."""
+
+    @abstractmethod
+    def get_texts_embedding(self, texts: List[str]) -> List[Embedding]:
+        """Get text embeddings."""
+
+    @abstractmethod
+    def get_documents_embedding(self, documents: List[Document]) -> List[Document]:
+        """Get documents embeddings."""
+
+    @staticmethod
+    def similarity(
+        embedding1: Embedding,
+        embedding2: Embedding,
+        mode: SimilarityMode = SimilarityMode.COSINE,
+    ):
+        """Get embedding similarity."""
+        return similarity(embedding1, embedding2, mode)
+
+    def __call__(self, documents: List[Document]) -> List[Document]:
+        return self.get_documents_embedding(documents)

beekeeper/core/evaluation/__init__.py

@@ -0,0 +1,5 @@
+from beekeeper.core.evaluation.context_similarity import (
+    ContextSimilarityEvaluator,
+)
+
+__all__ = ["ContextSimilarityEvaluator"]

beekeeper/core/evaluation/context_similarity.py

@@ -0,0 +1,73 @@
+from typing import Dict, List
+
+import numpy as np
+from beekeeper.core.embeddings import BaseEmbedding, SimilarityMode
+from pydantic.v1 import BaseModel
+
+
+class ContextSimilarityEvaluator(BaseModel):
+    """
+    Measures how much context has contributed to the answer’s.
+    A higher value suggests a greater proportion of the context is present in the LLM's response.
+
+    Args:
+        embed_model (BaseEmbedding): The embedding model used to compute vector representations.
+        similarity_mode (str, optional): Similarity strategy to use. Supported options are
+            `"cosine"`, `"dot_product"`, and `"euclidean"`. Defaults to `"cosine"`.
+        similarity_threshold (float, optional): Embedding similarity threshold for determining
+            whether a context segment "passes". Defaults to `0.8`.
+
+    Example:
+        .. code-block:: python
+
+            from beekeeper.core.evaluation import ContextSimilarityEvaluator
+            from beekeeper.embeddings.huggingface import HuggingFaceEmbedding
+
+            embedding = HuggingFaceEmbedding()
+            ctx_sim_evaluator = ContextSimilarityEvaluator(embed_model=embedding)
+    """
+
+    embed_model: BaseEmbedding
+    similarity_mode: SimilarityMode = SimilarityMode.COSINE
+    similarity_threshold: float = 0.8
+
+    class Config:
+        arbitrary_types_allowed = True
+
+    def evaluate(self, contexts: List[str], generated_text: str) -> Dict:
+        """
+        Args:
+            contexts (List[str]): List contexts used to generate LLM response.
+            generated_text (str): LLM response based on given context.
+
+        Example:
+            .. code-block:: python
+
+                evaluation_result = ctx_sim_evaluator.evaluate(
+                    contexts=[], generated_text="<candidate>"
+                )
+        """
+        if not contexts or not generated_text:
+            raise ValueError(
+                "Must provide these parameters [`contexts`, `generated_text`]",
+            )
+
+        evaluation_result = {"contexts_score": [], "score": 0}
+        candidate_embedding = self.embed_model.get_text_embedding(generated_text)
+
+        for context in contexts:
+            context_embedding = self.embed_model.get_text_embedding(context)
+            evaluation_result["contexts_score"].append(
+                self.embed_model.similarity(
+                    candidate_embedding,
+                    context_embedding,
+                    mode=self.similarity_mode,
+                ),
+            )
+
+        evaluation_result["score"] = np.mean(evaluation_result["contexts_score"])
+        evaluation_result["passing"] = (
+            evaluation_result["score"] >= self.similarity_threshold
+        )
+
+        return evaluation_result

beekeeper/core/flows/__init__.py

@@ -0,0 +1,3 @@
+from beekeeper.core.flows.ingestion_flow import IngestionFlow
+
+__all__ = ["IngestionFlow"]

beekeeper/core/flows/ingestion_flow.py

@@ -0,0 +1,178 @@
+from enum import Enum
+from typing import List, Optional
+
+from beekeeper.core.document import Document
+from beekeeper.core.readers import BaseReader
+from beekeeper.core.schema import TransformerComponent
+from beekeeper.core.vector_stores import BaseVectorStore
+
+
+class DocStrategy(Enum):
+    """
+    Document de-duplication strategies work by comparing the hashes in the vector store.
+    They require a vector store to be set.
+    """
+
+    DUPLICATE_ONLY = "duplicate_only"
+    DUPLICATE_AND_DELETE = "duplicate_and_delete"
+    DEDUPLICATE_OFF = "deduplicate_off"
+
+
+class IngestionFlow:
+    """
+    An ingestion flow for processing and storing data.
+
+    Args:
+        transformers (List[TransformerComponent]): A list of transformer components applied to the input documents.
+        doc_strategy (DocStrategy): The strategy used for handling document duplicates.
+            Defaults to `DocStrategy.DUPLICATE_ONLY`.
+        post_transformer (bool): Whether document de-duplication should be applied after transformation step.
+            Defaults to `False`.
+        readers (BaseReader, optional): List of readers for loading or fetching documents.
+        vector_store (BaseVectorStore, optional): Vector store for saving processed documents
+
+    Example:
+        .. code-block:: python
+
+            from beekeeper.core.flows import IngestionFlow
+            from beekeeper.core.text_chunkers import TokenTextChunker
+            from beekeeper.embeddings.huggingface import HuggingFaceEmbedding
+
+            ingestion_flow = IngestionFlow(
+                transformers=[
+                    TokenTextChunker(),
+                    HuggingFaceEmbedding(model_name="intfloat/multilingual-e5-small"),
+                ]
+            )
+    """
+
+    def __init__(
+        self,
+        transformers: List[TransformerComponent],
+        doc_strategy: DocStrategy = DocStrategy.DUPLICATE_ONLY,
+        post_transformer: bool = False,
+        readers: Optional[List[BaseReader]] = None,
+        vector_store: Optional[BaseVectorStore] = None,
+    ) -> None:
+        self.doc_strategy = doc_strategy
+        self.post_transformer = post_transformer
+        self.transformers = transformers
+        self.readers = readers
+        self.vector_store = vector_store
+
+    def _read_documents(self, documents: Optional[List[Document]]):
+        input_documents = []
+
+        if documents is not None:
+            input_documents.extend(documents)
+
+        if self.readers is not None:
+            for reader in self.readers:
+                input_documents.extend(reader.load_data())
+
+        return input_documents
+
+    def _handle_duplicates(self, documents) -> List[Document]:
+        ids, existing_hashes, existing_ref_hashes = (
+            self.vector_store.get_all_document_hashes()
+        )
+
+        if self.post_transformer:
+            # Use own document hash (chunks level) for de-duplication
+            hashes_fallback = existing_hashes
+        else:
+            # Use parent document hash `ref_doc_hash`
+            # Fallback to document own hash if `ref_doc_hash` (parent level) is missing for de-duplication
+            hashes_fallback = [
+                existing_ref_hashes[i]
+                if existing_ref_hashes[i] is not None
+                else existing_hashes[i]
+                for i in range(len(existing_ref_hashes))
+            ]
+
+        current_hashes = []
+        current_unique_hashes = []
+        dedup_documents_to_run = []
+
+        for doc in documents:
+            current_hashes.append(doc.hash)
+
+            if (
+                doc.hash not in hashes_fallback
+                and doc.hash not in current_unique_hashes
+                and doc.get_content() != ""
+            ):
+                dedup_documents_to_run.append(doc)
+                current_unique_hashes.append(
+                    doc.hash,
+                )  # Prevent duplicating same document hash in same batch flow execution.
+
+        if self.doc_strategy == DocStrategy.DUPLICATE_AND_DELETE:
+            ids_to_remove = [
+                ids[i]
+                for i in range(len(hashes_fallback))
+                if hashes_fallback[i] not in current_hashes
+            ]
+
+            if self.vector_store is not None:
+                self.vector_store.delete_documents(ids_to_remove)
+
+        return dedup_documents_to_run
+
+    def _run_transformers(
+        self,
+        documents: List[Document],
+        transformers: TransformerComponent,
+    ) -> List[Document]:
+        _documents = documents.copy()
+
+        for transformer in transformers:
+            _documents = transformer(_documents)
+
+        return _documents
+
+    def run(self, documents: List[Document] = None) -> List[Document]:
+        """
+        Run an ingestion flow.
+
+        Args:
+            documents: Set of documents to be transformed.
+
+        Example:
+            .. code-block:: python
+
+                ingestion_flow.run(documents: List[Document])
+        """
+        documents_processed = []
+        input_documents = self._read_documents(documents)
+
+        if (
+            self.vector_store is not None
+            and self.doc_strategy != DocStrategy.DEDUPLICATE_OFF
+            and not self.post_transformer
+        ):
+            # Apply transformers before de-dup (parent level)
+
+            documents_to_run = self._handle_duplicates(input_documents)
+        else:
+            # Apply transformers after de-dup (chunk level)
+            documents_to_run = input_documents
+
+        if documents_to_run:
+            documents_processed = self._run_transformers(
+                documents_to_run,
+                self.transformers,
+            )
+
+            # Apply transformers after de-dup (chunk level)
+            if (
+                self.vector_store is not None
+                and self.doc_strategy != DocStrategy.DEDUPLICATE_OFF
+                and self.post_transformer
+            ):
+                documents_processed = self._handle_duplicates(documents_processed)
+
+            if self.vector_store is not None and documents_processed:
+                self.vector_store.add_documents(documents_processed)
+
+        return documents_processed

beekeeper/core/llms/__init__.py

@@ -0,0 +1,11 @@
+from beekeeper.core.llms.base import BaseLLM
+from beekeeper.core.llms.types import (
+    ChatMessage,
+    ChatResponse,
+    GenerateResponse,
+    MessageRole,
+)
+
+__all__ = (
+    ["BaseLLM", "ChatMessage", "ChatResponse", "GenerateResponse", "MessageRole"],
+)

beekeeper/core/llms/base.py

@@ -0,0 +1,35 @@
+from abc import ABC, abstractmethod
+from typing import Any, List, Optional
+
+from beekeeper.core.llms.types import ChatMessage, ChatResponse, GenerateResponse
+from beekeeper.core.observers import BaseObserver
+from pydantic import BaseModel
+
+
+class BaseLLM(ABC, BaseModel):
+    """An interface for LLMs."""
+
+    model_config = {"arbitrary_types_allowed": True}
+    callback_manager: Optional[BaseObserver] = None
+
+    @classmethod
+    def class_name(cls) -> str:
+        return "BaseLLM"
+
+    def convert_chat_messages(self, messages: List[ChatMessage]) -> List[dict]:
+        """Convert chat messages to LLM message format."""
+        return [message.model_dump() for message in messages]
+
+    @abstractmethod
+    def completion(self, prompt: str, **kwargs: Any) -> GenerateResponse:
+        """Generates a completion for LLM."""
+
+    @abstractmethod
+    def text_completion(self, prompt: str, **kwargs: Any) -> str:
+        """Generates a text completion for LLM."""
+
+    @abstractmethod
+    def chat_completion(
+        self, messages: List[ChatMessage], **kwargs: Any
+    ) -> ChatResponse:
+        """Generates a chat completion for LLM."""

beekeeper/core/llms/decorators.py

@@ -0,0 +1,83 @@
+import asyncio
+import threading
+import time
+from logging import getLogger
+from typing import Callable
+
+from beekeeper.core.llms.types import ChatMessage
+from beekeeper.core.observers.types import PayloadRecord
+
+logger = getLogger(__name__)
+
+
+def llm_chat_observer() -> Callable:
+    """
+    Decorator to wrap a method with llm handler logic.
+    Looks for observability instances in `self.callback_manager`.
+    """
+
+    def decorator(f: Callable) -> Callable:
+        def async_wrapper(self, *args, **kwargs):
+            callback_manager_fns = getattr(self, "callback_manager", None)
+
+            start_time = time.time()
+            llm_return_val = f(self, *args, **kwargs)
+            response_time = int((time.time() - start_time) * 1000)
+
+            if callback_manager_fns:
+
+                def async_callback_thread():
+                    try:
+                        # Extract input messages
+                        if len(args) > 0 and isinstance(args[0], ChatMessage):
+                            input_chat_messages = args[0]
+                        elif "messages" in kwargs:
+                            input_chat_messages = kwargs["messages"]
+                        else:
+                            raise ValueError(
+                                "No messages provided in positional or keyword arguments"
+                            )
+
+                        # Get the user's latest message after each interaction to chat observability.
+                        user_messages = [
+                            msg for msg in input_chat_messages if msg.role == "user"
+                        ]
+                        last_user_message = (
+                            user_messages[-1].content if user_messages else None
+                        )
+
+                        # Get the system/instruct (first) message to chat observability.
+                        system_messages = [
+                            msg for msg in input_chat_messages if msg.role == "system"
+                        ]
+                        system_message = (
+                            system_messages[0].content if system_messages else None
+                        )
+
+                        callback = callback_manager_fns(
+                            payload=PayloadRecord(
+                                input_text=(system_message or "") + last_user_message,
+                                generated_text=llm_return_val.message.content,
+                                generated_token_count=llm_return_val.raw["usage"][
+                                    "completion_tokens"
+                                ],
+                                input_token_count=llm_return_val.raw["usage"][
+                                    "prompt_tokens"
+                                ],
+                                response_time=response_time,
+                            )
+                        )
+
+                        if asyncio.iscoroutine(callback):
+                            asyncio.run(callback)
+
+                    except Exception as e:
+                        logger.error(f"Observability callback error: {e}")
+
+                threading.Thread(target=async_callback_thread).start()
+
+            return llm_return_val
+
+        return async_wrapper
+
+    return decorator

beekeeper/core/llms/types.py

@@ -0,0 +1,33 @@
+from enum import Enum
+from typing import Any, Optional
+
+from pydantic import BaseModel, Field
+
+
+class MessageRole(str, Enum):
+    ASSISTANT = "assistant"
+    SYSTEM = "system"
+    USER = "user"
+    TOOL = "tool"
+
+
+class ChatMessage(BaseModel):
+    """Chat message."""
+
+    model_config = {"use_enum_values": True}
+    role: MessageRole = Field(default=MessageRole.USER)
+    content: Optional[str] = Field(default=None)
+
+
+class GenerateResponse(BaseModel):
+    """Generate response."""
+
+    text: str = Field(..., description="Generated text response")
+    raw: Optional[Any] = Field(default=None)
+
+
+class ChatResponse(BaseModel):
+    """Chat completion response."""
+
+    message: ChatMessage
+    raw: Optional[Any] = None

beekeeper/core/observers/__init__.py

@@ -0,0 +1,3 @@
+from beekeeper.core.observers.base import BaseObserver, ModelObserver
+
+__all__ = (["BaseObserver", "ModelObserver"],)

beekeeper/core/observers/base.py

@@ -0,0 +1,36 @@
+from abc import ABC, abstractmethod
+from typing import Optional
+
+from beekeeper.core.observers.types import PayloadRecord
+from beekeeper.core.prompts import PromptTemplate
+
+
+class BaseObserver(ABC):
+    """An interface for observability."""
+
+    @classmethod
+    def class_name(cls) -> str:
+        return "BaseObserver"
+
+
+class ModelObserver(BaseObserver):
+    """An interface for model observability."""
+
+    def __init__(self, prompt_template: Optional[PromptTemplate] = None) -> None:
+        self.prompt_template = prompt_template
+
+    @classmethod
+    def class_name(cls) -> str:
+        return "ModelObserver"
+
+    @abstractmethod
+    def __call__(self, payload: PayloadRecord) -> None:
+        """ModelObserver."""
+
+
+class TelemetryObserver(BaseObserver):
+    """An interface for telemetry observability."""
+
+    @classmethod
+    def class_name(cls) -> str:
+        return "TelemetryObserver"

beekeeper/core/observers/types.py

@@ -0,0 +1,11 @@
+from pydantic import BaseModel
+
+
+class PayloadRecord(BaseModel):
+    """Payload record."""
+
+    input_text: str
+    generated_text: str
+    generated_token_count: int
+    input_token_count: int
+    response_time: int

beekeeper/core/prompts/__init__.py

@@ -0,0 +1,3 @@
+from beekeeper.core.prompts.base import PromptTemplate
+
+__all__ = (["PromptTemplate"],)