配置文档摘要能力(Python SDK)

基于已有的知识库，进行摘要总结。有stuff、refine、map-reduce策略。

• Stuff：将所有文档直接填充到prompt中，提给模型处理，适合文档较少的场景。

  from pangukitsappdev.api.embeddings.factory import Embeddings
  from pangukitsappdev.api.llms.factory import LLMs
  from pangukitsappdev.api.memory.vector.factory import Vectors
  from pangukitsappdev.api.memory.vector.vector_config import VectorStoreConfig, ServerInfoCss
  from pangukitsappdev.skill.doc.summary import DocSummaryStuffSkill
  vector_store_config = VectorStoreConfig(store_name="css",
                                          index_name="your_index_name",
                                          embedding=Embeddings.of("css"),
                                          text_key="name",
                                          vector_fields=["description"],
                                          distance_strategy="inner_product",
                                          server_info=ServerInfoCss(env_prefix="sdk.memory.css"))
  vector_api = Vectors.of("css", vector_store_config)
  
  # 检索
  query = "杜甫"
  docs = vector_api.similarity_search(query, 4)
  
  # 摘要
  doc_skill = DocSummaryStuffSkill(LLMs.of("pangu"))
  
  print(doc_skill.execute({"documents": docs}))

• Refine：基于首个文档，并循环后续文档来迭代更新答案。

  from pangukitsappdev.api.embeddings.factory import Embeddings
  from pangukitsappdev.api.llms.factory import LLMs
  from pangukitsappdev.api.memory.vector.factory import Vectors
  from pangukitsappdev.api.memory.vector.vector_config import VectorStoreConfig, ServerInfoCss
  from pangukitsappdev.skill.doc.summary import DocSummaryRefineSkill
  vector_store_config = VectorStoreConfig(store_name="css",
                                          index_name="your_index_name",
                                          embedding=Embeddings.of("css"),
                                          text_key="name",
                                          vector_fields=["description"],
                                          distance_strategy="inner_product",
                                          server_info=ServerInfoCss(env_prefix="sdk.memory.css"))
  
  vector_api = Vectors.of("css", vector_store_config)
  
  # 检索
  query = "杜甫"
  docs = vector_api.similarity_search(query, 4)
  
  # 摘要
  doc_skill = DocSummaryRefineSkill(LLMs.of("pangu"))
  
  print(doc_skill.execute({"documents": docs}))

• Map-Reduce：先将文档单独进行摘要， 将摘要后的文档再提交给模型。 必要时，会循环迭代摘要。

  from pangukitsappdev.api.embeddings.factory import Embeddings
  from pangukitsappdev.api.llms.factory import LLMs
  from pangukitsappdev.api.memory.vector.factory import Vectors
  from pangukitsappdev.api.memory.vector.vector_config import VectorStoreConfig, ServerInfoCss
  from pangukitsappdev.skill.doc.summary import DocSummaryMapReduceSkill
  vector_store_config = VectorStoreConfig(store_name="css",
                                          index_name="your_index_name",
                                          embedding=Embeddings.of("css"),
                                          text_key="name",
                                          vector_fields=["description"],
                                          distance_strategy="inner_product",
                                          server_info=ServerInfoCss(env_prefix="sdk.memory.css"))
  
  vector_api = Vectors.of("css", vector_store_config)
  
  # 检索
  query = "杜甫"
  docs = vector_api.similarity_search(query, 4)
  
  # 摘要
  doc_skill = DocSummaryMapReduceSkill(LLMs.of("pangu"))
  
  print(doc_skill.execute({"documents": docs}))