ragwiki-corpusgraph

Description

A corpus graph for Wikipedia corpus (rag:nq_wiki), built using ragwiki-terrier sparse index.
The graph encodes semantic relationships between documents by connecting each document to its k=16 nearest neighbors based on bm25 retriever.

Usage

# Load the artifact
import pyterrier as pt
artifact = pt.Artifact.from_hf('astappiev/ragwiki-corpusgraph')

Usage with GAR

import pyterrier as pt
from pyterrier_adaptive import GAR, NpTopKCorpusGraph
from pyterrier_t5 import MonoT5ReRanker

sparse_index: pt.terrier.TerrierIndex = pt.Artifact.from_hf("pyterrier/ragwiki-terrier")
retriever = pt.rewrite.tokenise() >> sparse_index.bm25(include_fields=["docno", "text", "title"]) >> pt.rewrite.reset()

get_text = sparse_index.text_loader(["docno", "text", "title"])
prepare_text = pt.apply.generic(lambda df: df.assign(qid=df["qid"].map(str), docno=df["docno"].map(str))) >> get_text
scorer = prepare_text >> MonoT5ReRanker(verbose=False, batch_size=64)

graph: NpTopKCorpusGraph = pt.Artifact.from_hf("astappiev/ragwiki-corpusgraph").to_limit_k(8)

pipeline = retriever >> GAR(scorer, graph) >> get_text
pipeline.search("hello world")

Benchmarks

TODO: Provide benchmarks for the artifact.

Reproduction

This graph was constructed using PyTerrier Adaptive.

import pyterrier as pt
from pyterrier_adaptive import NpTopKCorpusGraph

dataset: pt.datasets.Dataset = pt.get_dataset("rag:nq_wiki")
sparse_index: pt.terrier.TerrierIndex = pt.Artifact.from_hf("pyterrier/ragwiki-terrier")
bm25 = pt.rewrite.tokenise() >> sparse_index.bm25(include_fields=["docno", "text", "title"], threads=slurm_cpus) >> pt.rewrite.reset()

graph = NpTopKCorpusGraph.from_retriever(
    bm25 % (build_graph_k + 1),
    dataset.get_corpus_iter(),
    "../index/ragwiki-corpusgraph",
    k=build_graph_k,
    batch_size=65_536,
)
graph.to_hf('astappiev/ragwiki-corpusgraph')

However the code above will take 2 months to run on a single machine. The graph was built using modified version of from_retriever to enable parallel compute on the cluster.

It took around 14340 CPU-hours to build the graph on our cluster.

Metadata

{
  "type": "corpus_graph",
  "format": "np_topk",
  "package_hint": "pyterrier-adaptive",
  "doc_count": 21015324,
  "k": 16
}