Recency Filtering

Showcase capabilities of recency-weighted node postprocessor

from llama_index import GPTVectorStoreIndex, SimpleDirectoryReader, ServiceContext
from llama_index.indices.postprocessor import (
    FixedRecencyPostprocessor,
    EmbeddingRecencyPostprocessor
)
from llama_index.node_parser import SimpleNodeParser
from llama_index.storage.docstore import SimpleDocumentStore
from llama_index.response.notebook_utils import display_response

/Users/jerryliu/Programming/llama_index/.venv/lib/python3.10/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

Parse Documents into Nodes, add to Docstore

In this example, there are 3 different versions of PG’s essay. They are largely identical except for one specific section, which details the amount of funding they raised for Viaweb.

V1: 50k, V2: 30k, V3: 10K

V1: 2020-01-01, V2: 2020-02-03, V3: 2022-04-12

The idea is to encourage index to fetch the most recent info (which is V3)

# load documents
from llama_index.storage.storage_context import StorageContext


def get_file_metadata(file_name: str):
    """Get file metadata."""
    if "v1" in file_name:
        return {"date": "2020-01-01"}
    elif "v2" in file_name:
        return {"date": "2020-02-03"}
    elif "v3" in file_name:
        return {"date": "2022-04-12"}
    else:
        raise ValueError("invalid file")

documents = SimpleDirectoryReader(
    input_files=[
        'test_versioned_data/paul_graham_essay_v1.txt',
        'test_versioned_data/paul_graham_essay_v2.txt',
        'test_versioned_data/paul_graham_essay_v3.txt'
    ],
    file_metadata=get_file_metadata
).load_data()

# define service context (wrapper container around current classes)
service_context = ServiceContext.from_defaults(chunk_size=512)

# use node parser in service context to parse into nodes
nodes = service_context.node_parser.get_nodes_from_documents(documents)

# add to docstore
docstore = SimpleDocumentStore()
docstore.add_documents(nodes)

storage_context = StorageContext.from_defaults(docstore=docstore)

print(documents[2].get_text())

Build Index

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

INFO:llama_index.token_counter.token_counter:> [build_index_from_nodes] Total LLM token usage: 0 tokens
INFO:llama_index.token_counter.token_counter:> [build_index_from_nodes] Total embedding token usage: 84471 tokens

Define Recency Postprocessors

node_postprocessor = FixedRecencyPostprocessor(service_context=service_context)

node_postprocessor_emb = EmbeddingRecencyPostprocessor(service_context=service_context)

Query Index

# 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?", 
)

INFO:llama_index.token_counter.token_counter:> [query] Total LLM token usage: 1813 tokens
INFO:llama_index.token_counter.token_counter:> [query] Total embedding token usage: 22 tokens

# query using fixed recency 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 embedding-based 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?", 
)

INFO:llama_index.token_counter.token_counter:> [query] Total LLM token usage: 541 tokens
INFO:llama_index.token_counter.token_counter:> [query] Total embedding token usage: 22 tokens

Query Index (Lower-Level Usage)

In this example we first get the full set of nodes from a query call, and then send to node postprocessor, and then finally synthesize response through a list index.

from llama_index import GPTListIndex

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

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

INFO:llama_index.token_counter.token_counter:> [query] Total LLM token usage: 0 tokens
INFO:llama_index.token_counter.token_counter:> [query] Total embedding token usage: 22 tokens

list_index = GPTListIndex(resp_nodes)
query_engine = list_index.as_query_engine(
    node_postprocessors=[node_postprocessor]
)
response = query_engine.query(query_str)

INFO:llama_index.token_counter.token_counter:> [build_index_from_nodes] Total LLM token usage: 0 tokens
INFO:llama_index.token_counter.token_counter:> [build_index_from_nodes] Total embedding token usage: 0 tokens
INFO:llama_index.token_counter.token_counter:> [query] Total LLM token usage: 541 tokens
INFO:llama_index.token_counter.token_counter:> [query] Total embedding token usage: 0 tokens