vruksha

vruksha turns a litesearch store into a typed knowledge graph. It extracts entities and typed relations from each chunk with a model you pass in, primes the model with the entities already in the graph so it reuses them, merges the duplicates that slip through behind a lexical guard, and searches by walking refers_to/cites/builds_on edges fused with hybrid.

It is model-agnostic: you inject chat (any object with .oneshot(prompt, sp=, max_tokens=)) and emb_fn. rishi, urai, an API client, anything. vruksha names no model and owns litesearch’s entity/mention/edge tables.

from litesearch import database
from vruksha import build_graph          # extract + prime + write + embed + merge, incremental

db = database('corpus.db')               # a litesearch store with a tree
build_graph(db, chat=my_chat, emb_fn=my_encoder)   # a local model does the extraction
db.graph_search('how does resolution work', qemb)  # hybrid + a typed-edge walk

The lexical guard

Priming stops most duplicates; resolve_entities merges the rest. Embedding similarity alone merges python 3.11 into python 3.12, so every merge passes _lex_ok first: token overlap or a matching acronym or, for a one-word name, the same romanisation key (Bhīṣma, Bheeshma, भीष्म), and identical numbers.

from vruksha.entities import _lex_ok
_lex_ok('multi-head attention', 'multi head attention'), _lex_ok('python 3.11', 'python 3.12')

A per-corpus citation seed

The model catches loose references. Structured ones (an EU directive’s Article 5, Annex III) are cheaper caught by a regex, so a legal corpus can pass seed_fn=cites, canon=norm_cite. It is optional and off by default: nothing but a citation-dense corpus needs it.

What it is worth

On cross-reference bridges (the query shares no word with the answer), the typed leg reaches the target 0.61 to 0.87 of the time against 0.08 to 0.20 for hybrid, tracking the extractor: a local Qwen3-4B matches a frontier model, a 1.5B needs the seed. The PMI co-occurrence graph this package used to build reached none of them; the method and numbers are in litesearch evals/RESULTS.md.

Install

pip install vruksha