vruksha

Four problems, and none of them needs a model.

from litesearch import database
from vruksha import build_graph, resolve_entities

db = database('corpus.db')
build_graph(db, rows, emb_fn=enc)     # entities and co-occurrence edges
resolve_entities(db)                  # `hnsw` and `HNSW index` become one node
db.graph_search('how does resolution work', qemb, graph_w=0.5)

The lexical guard

Embedding similarity alone merges python 3.11 into python 3.12. _lex_ok requires token overlap, matching digits and a matching acronym before a merge goes through.

from vruksha.entities import _lex_ok

_lex_ok('usearch', 'usearch index'), _lex_ok('python 3.11', 'python 3.12')

When to turn the search leg on

Measured against plain hybrid search:

• Regulation and legal text: a loss. p_mrr 0.8170 for plain hybrid against 0.7395, 0.6859 and 0.6463 at graph_w 0.25, 0.5 and 1.0, at two to four times the latency.
• Papers and prose: a win. Better in seven of nine paired-bootstrap comparisons, +0.0387 target MRR on arXiv at graph_w=1.0.

So graph_search is opt-in by name and off by default. Turn it on for a corpus whose entities carry meaning, and raise graph_w towards 1.0 when you do.

Install

pip install vruksha