v0.1.0 — feature-complete pre-alpha

The first feature-complete pre-alpha. Multi-component scoring, entity-linked
KG mode, three adapters (InMemory, Qdrant, Chroma), reasoning-path
explanation, document chunking, latency optimisations, and an
LLM-based KG construction option.

Headline benchmark

HotpotQA dev N=500, supporting-fact recall@10:

• naive cosine top-k: 81.9%
• hubmesh (multi-component over entity KG): 85.6% (+3.7 pts)
• HippoRAG-style PPR-only on the same KG: 56.5%

MuSiQue dev N=300 — win grows with hop count: +1.7 / +1.9 / +2.8 pts at
2-/3-/4-hop. Per-query latency ~22 ms mean / 26 ms p95 on a 7K-node KG.

See BENCHMARKS.md
for methodology, ablations, and caveats.

What's new in 0.1.0

• Reasoning paths — RetrievalResult.reasoning now contains multi-hop
  traces from query entities to retrieved documents.
• Embedding-based entity linker — clusters surface variations
  ("U.S." / "USA" / "United States") via sentence-transformer cosine.
• Chroma adapter (ephemeral / persistent / HTTP modes).
• LLM-based KG construction — provider-agnostic, with passage-hash caching.
• Document chunking helpers for splitting long source documents.
• PPRSolver — precomputes the sparse transition matrix; ~10× faster
  per-query PPR.
• Vectorised document scoring in KG mode.

Install

pip install hubmesh                 # core
pip install "hubmesh[all]"          # everything
python -m spacy download en_core_web_sm