Source code for langchain_community.retrievers.svm

from __future__ import annotations

import concurrent.futures
from typing import Any, Iterable, List, Optional

import numpy as np
from langchain_core.callbacks import CallbackManagerForRetrieverRun
from langchain_core.documents import Document
from langchain_core.embeddings import Embeddings
from langchain_core.retrievers import BaseRetriever


def create_index(contexts: List[str], embeddings: Embeddings) -> np.ndarray:
    """为上下文列表创建嵌入索引。

参数：
    contexts：要嵌入的上下文列表。
    embeddings：要使用的嵌入模型。

返回：
    嵌入的索引。
"""
    with concurrent.futures.ThreadPoolExecutor() as executor:
        return np.array(list(executor.map(embeddings.embed_query, contexts)))


class SVMRetriever(BaseRetriever):
    """`SVM` 检索器。

    主要基于
    https://github.com/karpathy/randomfun/blob/master/knn_vs_svm.ipynb"""

    embeddings: Embeddings
    """要使用的嵌入模型。"""
    index: Any
    """索引嵌入。"""
    texts: List[str]
    """要索引的文本列表。"""
    metadatas: Optional[List[dict]] = None
    """每个文本对应的元数据列表。"""
    k: int = 4
    """返回结果的数量。"""
    relevancy_threshold: Optional[float] = None
    """相关性的阈值。"""

    class Config:

        """此pydantic对象的配置。"""

        arbitrary_types_allowed = True

    @classmethod
    def from_texts(
        cls,
        texts: List[str],
        embeddings: Embeddings,
        metadatas: Optional[List[dict]] = None,
        **kwargs: Any,
    ) -> SVMRetriever:
        index = create_index(texts, embeddings)
        return cls(
            embeddings=embeddings,
            index=index,
            texts=texts,
            metadatas=metadatas,
            **kwargs,
        )

    @classmethod
    def from_documents(
        cls,
        documents: Iterable[Document],
        embeddings: Embeddings,
        **kwargs: Any,
    ) -> SVMRetriever:
        texts, metadatas = zip(*((d.page_content, d.metadata) for d in documents))
        return cls.from_texts(
            texts=texts, embeddings=embeddings, metadatas=metadatas, **kwargs
        )

    def _get_relevant_documents(
        self, query: str, *, run_manager: CallbackManagerForRetrieverRun
    ) -> List[Document]:
        try:
            from sklearn import svm
        except ImportError:
            raise ImportError(
                "Could not import scikit-learn, please install with `pip install "
                "scikit-learn`."
            )

        query_embeds = np.array(self.embeddings.embed_query(query))
        x = np.concatenate([query_embeds[None, ...], self.index])
        y = np.zeros(x.shape[0])
        y[0] = 1

        clf = svm.LinearSVC(
            class_weight="balanced", verbose=False, max_iter=10000, tol=1e-6, C=0.1
        )
        clf.fit(x, y)

        similarities = clf.decision_function(x)
        sorted_ix = np.argsort(-similarities)

        # svm.LinearSVC in scikit-learn is non-deterministic.
        # if a text is the same as a query, there is no guarantee
        # the query will be in the first index.
        # this performs a simple swap, this works because anything
        # left of the 0 should be equivalent.
        zero_index = np.where(sorted_ix == 0)[0][0]
        if zero_index != 0:
            sorted_ix[0], sorted_ix[zero_index] = sorted_ix[zero_index], sorted_ix[0]

        denominator = np.max(similarities) - np.min(similarities) + 1e-6
        normalized_similarities = (similarities - np.min(similarities)) / denominator

        top_k_results = []
        for row in sorted_ix[1 : self.k + 1]:
            if (
                self.relevancy_threshold is None
                or normalized_similarities[row] >= self.relevancy_threshold
            ):
                metadata = self.metadatas[row - 1] if self.metadatas else {}
                doc = Document(page_content=self.texts[row - 1], metadata=metadata)
                top_k_results.append(doc)
        return top_k_results