时间加权重排序¶

展示时间加权节点后处理器的功能

In [ ]:

from llama_index.core import VectorStoreIndex, SimpleDirectoryReader
from llama_index.core.postprocessor import TimeWeightedPostprocessor
from llama_index.core.node_parser import SentenceSplitter
from llama_index.core.storage.docstore import SimpleDocumentStore
from llama_index.core.response.notebook_utils import display_response
from datetime import datetime, timedelta

from llama_index.core import VectorStoreIndex, SimpleDirectoryReader from llama_index.core.postprocessor import TimeWeightedPostprocessor from llama_index.core.node_parser import SentenceSplitter from llama_index.core.storage.docstore import SimpleDocumentStore from llama_index.core.response.notebook_utils import display_response from datetime import datetime, timedelta

/home/loganm/miniconda3/envs/llama-index/lib/python3.11/site-packages/tqdm/auto.py:21: TqdmWarning: IProgress not found. Please update jupyter and ipywidgets. See https://ipywidgets.readthedocs.io/en/stable/user_install.html
  from .autonotebook import tqdm as notebook_tqdm

将文档解析为节点并存入文档库¶

在本示例中，存在 PG 文章的 3 个不同版本。这些版本大体相同，但有一个特定部分存在差异——该部分详细说明了他们为 Viaweb 筹集的资金金额。

V1 版本：5 万美元
V2 版本：3 万美元
V3 版本：1 万美元

V1 版本：-1 天（旧）
V2 版本：-2 天（更旧）
V3 版本：-3 天（最旧）

设计目的是让索引优先获取最新信息（即 V3 版本）

In [ ]:

# load documents
from llama_index.core import StorageContext


now = datetime.now()
key = "__last_accessed__"


doc1 = SimpleDirectoryReader(
    input_files=["./test_versioned_data/paul_graham_essay_v1.txt"]
).load_data()[0]


doc2 = SimpleDirectoryReader(
    input_files=["./test_versioned_data/paul_graham_essay_v2.txt"]
).load_data()[0]

doc3 = SimpleDirectoryReader(
    input_files=["./test_versioned_data/paul_graham_essay_v3.txt"]
).load_data()[0]


# define settings
from llama_index.core import Settings

Settings.text_splitter = SentenceSplitter(chunk_size=512)

# use node parser from settings to parse docs into nodes
nodes1 = Settings.text_splitter.get_nodes_from_documents([doc1])
nodes2 = Settings.text_splitter.get_nodes_from_documents([doc2])
nodes3 = Settings.text_splitter.get_nodes_from_documents([doc3])


# fetch the modified chunk from each document, set metadata
# also exclude the date from being read by the LLM
nodes1[14].metadata[key] = (now - timedelta(hours=3)).timestamp()
nodes1[14].excluded_llm_metadata_keys = [key]

nodes2[14].metadata[key] = (now - timedelta(hours=2)).timestamp()
nodes2[14].excluded_llm_metadata_keys = [key]

nodes3[14].metadata[key] = (now - timedelta(hours=1)).timestamp()
nodes2[14].excluded_llm_metadata_keys = [key]


# add to docstore
docstore = SimpleDocumentStore()
nodes = [nodes1[14], nodes2[14], nodes3[14]]
docstore.add_documents(nodes)

storage_context = StorageContext.from_defaults(docstore=docstore)

# load documents from llama_index.core import StorageContext now = datetime.now() key = "__last_accessed__" doc1 = SimpleDirectoryReader( input_files=["./test_versioned_data/paul_graham_essay_v1.txt"] ).load_data()[0] doc2 = SimpleDirectoryReader( input_files=["./test_versioned_data/paul_graham_essay_v2.txt"] ).load_data()[0] doc3 = SimpleDirectoryReader( input_files=["./test_versioned_data/paul_graham_essay_v3.txt"] ).load_data()[0] # define settings from llama_index.core import Settings Settings.text_splitter = SentenceSplitter(chunk_size=512) # use node parser from settings to parse docs into nodes nodes1 = Settings.text_splitter.get_nodes_from_documents([doc1]) nodes2 = Settings.text_splitter.get_nodes_from_documents([doc2]) nodes3 = Settings.text_splitter.get_nodes_from_documents([doc3]) # fetch the modified chunk from each document, set metadata # also exclude the date from being read by the LLM nodes1[14].metadata[key] = (now - timedelta(hours=3)).timestamp() nodes1[14].excluded_llm_metadata_keys = [key] nodes2[14].metadata[key] = (now - timedelta(hours=2)).timestamp() nodes2[14].excluded_llm_metadata_keys = [key] nodes3[14].metadata[key] = (now - timedelta(hours=1)).timestamp() nodes2[14].excluded_llm_metadata_keys = [key] # add to docstore docstore = SimpleDocumentStore() nodes = [nodes1[14], nodes2[14], nodes3[14]] docstore.add_documents(nodes) storage_context = StorageContext.from_defaults(docstore=docstore)

构建索引¶

In [ ]:

# build index
index = VectorStoreIndex(nodes, storage_context=storage_context)

# build index index = VectorStoreIndex(nodes, storage_context=storage_context)

定义时效性后处理器¶

In [ ]:

node_postprocessor = TimeWeightedPostprocessor(
    time_decay=0.5, time_access_refresh=False, top_k=1
)

node_postprocessor = TimeWeightedPostprocessor( time_decay=0.5, time_access_refresh=False, top_k=1 )

查询索引¶

In [ ]:

# naive query
query_engine = index.as_query_engine(
    similarity_top_k=3,
)
response = query_engine.query(
    "How much did the author raise in seed funding from Idelle's husband"
    " (Julian) for Viaweb?",
)

# naive query query_engine = index.as_query_engine( similarity_top_k=3, ) response = query_engine.query( "How much did the author raise in seed funding from Idelle's husband" " (Julian) for Viaweb?", )

In [ ]:

display_response(response)

display_response(response)

Final Response: $50,000

In [ ]:

# query using time weighted node postprocessor

query_engine = index.as_query_engine(
    similarity_top_k=3, node_postprocessors=[node_postprocessor]
)
response = query_engine.query(
    "How much did the author raise in seed funding from Idelle's husband"
    " (Julian) for Viaweb?",
)

# query using time weighted node postprocessor query_engine = index.as_query_engine( similarity_top_k=3, node_postprocessors=[node_postprocessor] ) response = query_engine.query( "How much did the author raise in seed funding from Idelle's husband" " (Julian) for Viaweb?", )

In [ ]:

display_response(response)

display_response(response)

Final Response: The author raised $10,000 in seed funding from Idelle's husband (Julian) for Viaweb.

查询索引（底层用法）¶

在本示例中，我们首先通过查询调用获取完整的节点集合，随后将其传递至节点后处理器，最终通过摘要索引合成响应。

In [ ]:

from llama_index.core import SummaryIndex

from llama_index.core import SummaryIndex

In [ ]:

query_str = (
    "How much did the author raise in seed funding from Idelle's husband"
    " (Julian) for Viaweb?"
)

query_str = ( "How much did the author raise in seed funding from Idelle's husband" " (Julian) for Viaweb?" )

In [ ]:

query_engine = index.as_query_engine(
    similarity_top_k=3, response_mode="no_text"
)
init_response = query_engine.query(
    query_str,
)
resp_nodes = [n for n in init_response.source_nodes]

query_engine = index.as_query_engine( similarity_top_k=3, response_mode="no_text" ) init_response = query_engine.query( query_str, ) resp_nodes = [n for n in init_response.source_nodes]

In [ ]:

# get the post-processed nodes -- which should be the top-1 sorted by date
new_resp_nodes = node_postprocessor.postprocess_nodes(resp_nodes)

summary_index = SummaryIndex([n.node for n in new_resp_nodes])
query_engine = summary_index.as_query_engine()
response = query_engine.query(query_str)

# get the post-processed nodes -- which should be the top-1 sorted by date new_resp_nodes = node_postprocessor.postprocess_nodes(resp_nodes) summary_index = SummaryIndex([n.node for n in new_resp_nodes]) query_engine = summary_index.as_query_engine() response = query_engine.query(query_str)

In [ ]:

display_response(response)

display_response(response)

Final Response: The author raised $10,000 in seed funding from Idelle's husband (Julian) for Viaweb.