Skip to content

Embedding recency

节点后处理器模块。

EmbeddingRecencyPostprocessor #

Bases: BaseNodePostprocessor

嵌入最近性后处理器。

Source code in llama_index/core/postprocessor/node_recency.py
 85
 86
 87
 88
 89
 90
 91
 92
 93
 94
 95
 96
 97
 98
 99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
class EmbeddingRecencyPostprocessor(BaseNodePostprocessor):
    """嵌入最近性后处理器。"""

    embed_model: BaseEmbedding = Field(default_factory=lambda: Settings.embed_model)
    date_key: str = "date"
    similarity_cutoff: float = Field(default=0.7)
    query_embedding_tmpl: str = Field(default=DEFAULT_QUERY_EMBEDDING_TMPL)

    @classmethod
    def class_name(cls) -> str:
        return "EmbeddingRecencyPostprocessor"

    def _postprocess_nodes(
        self,
        nodes: List[NodeWithScore],
        query_bundle: Optional[QueryBundle] = None,
    ) -> List[NodeWithScore]:
        """后处理节点。"""
        if query_bundle is None:
            raise ValueError("Missing query bundle in extra info.")

        # sort nodes by date
        node_dates = pd.to_datetime(
            [node.node.metadata[self.date_key] for node in nodes]
        )
        sorted_node_idxs = np.flip(node_dates.argsort())
        sorted_nodes: List[NodeWithScore] = [nodes[idx] for idx in sorted_node_idxs]

        # get embeddings for each node
        texts = [node.get_content(metadata_mode=MetadataMode.EMBED) for node in nodes]
        text_embeddings = self.embed_model.get_text_embedding_batch(texts=texts)

        node_ids_to_skip: Set[str] = set()
        for idx, node in enumerate(sorted_nodes):
            if node.node.node_id in node_ids_to_skip:
                continue
            # get query embedding for the "query" node
            # NOTE: not the same as the text embedding because
            # we want to optimize for retrieval results

            query_text = self.query_embedding_tmpl.format(
                context_str=node.node.get_content(metadata_mode=MetadataMode.EMBED),
            )
            query_embedding = self.embed_model.get_query_embedding(query_text)

            for idx2 in range(idx + 1, len(sorted_nodes)):
                if sorted_nodes[idx2].node.node_id in node_ids_to_skip:
                    continue
                node2 = sorted_nodes[idx2]
                if (
                    np.dot(query_embedding, text_embeddings[idx2])
                    > self.similarity_cutoff
                ):
                    node_ids_to_skip.add(node2.node.node_id)

        return [
            node for node in sorted_nodes if node.node.node_id not in node_ids_to_skip
        ]