hubmesh

hubmesh

Centrality-aware GraphRAG retrieval planner. Drop-in layer over any vector DB.

hubmesh is a Python library that improves multi-hop RAG quality on top of an existing vector database. You don't replace your infrastructure — you add a smart planner between your vector DB and your LLM.

What problem this solves

Naive vector retrieval ("embed query, get top-k by cosine similarity") fails on multi-hop questions like "Where was the founder of the company that acquired Slack born?" The correct answer requires retrieving entities along a reasoning path, not the single most similar item.

GraphRAG and HippoRAG showed that running a small Personalized PageRank over a knowledge graph at query time can substantially improve multi-hop retrieval. hubmesh extends that line with two contributions:

1. Multi-component seed selection. Instead of picking PPR seeds by raw query similarity (which picks wrong-community seeds at high feature overlap), seeds are chosen by a multi-component score combining query relevance, structural fit, and coverage diversity.
2. Budget-aware context packing. Once relevant entities are scored, pack them into the LLM's context window with explicit coverage and redundancy control rather than just truncating top-k.

The multi-component scoring pattern is adapted from the NNSI framework (Naidu Dsk, ICOMP'25 — to appear) for SDN topology optimization, repurposed here for retrieval planning.

Quickstart

In-memory (testing, small corpora)

from hubmesh import Planner
from hubmesh.adapters import InMemoryStore

embed = ...   # callable: text -> np.ndarray
docs = [...]  # list of Document or strings or dicts

store = InMemoryStore.from_documents(docs, embed=embed)
planner = Planner(store=store, embed=embed)
result = planner.retrieve(query="...", top_k=10, budget_tokens=4000)

Qdrant adapter (production)

from hubmesh import Planner
from hubmesh.adapters import QdrantStore

store = QdrantStore.from_documents(docs)                          # in-memory
store = QdrantStore.from_documents(docs, path="./qdrant_data")    # on-disk
store = QdrantStore.from_documents(docs, url="http://localhost:6333")  # remote

planner = Planner(store=store, embed=embed)
result = planner.retrieve(query="...", top_k=10)

Chroma adapter

from hubmesh.adapters import ChromaStore

store = ChromaStore.from_documents(docs)                          # ephemeral
store = ChromaStore.from_documents(docs, persist_directory="./chroma_data")
store = ChromaStore.from_documents(docs, host="localhost", port=8000)

Multi-hop / KG mode

from hubmesh.kg import build_entity_kg
import spacy

nlp = spacy.load("en_core_web_sm")
kg = build_entity_kg(docs, nlp=nlp)

planner = Planner(store=store, kg=kg, nlp=nlp)
result = planner.retrieve(query="Where was the founder of the company that bought Slack born?",
                          top_k=10, budget_tokens=4000)

# RetrievalResult includes reasoning paths showing why each doc was returned
for path in result.reasoning:
    print(f"  score={path.score:.3f}  {' → '.join(path.node_ids)}")

LLM-extracted KG (richer than spaCy)

from hubmesh.kg_llm import build_entity_kg_llm
from hubmesh.entity_linker import EmbeddingLinker, make_st_embedder

def llm(prompt):  # provider-agnostic — bring your own
    return your_llm_call(prompt)

kg = build_entity_kg_llm(docs, llm=llm, cache_path="kg_cache.json")

# optional: cross-document entity dedup — same Linker protocol as the spaCy path
kg = build_entity_kg_llm(docs, llm=llm, cache_path="kg_cache.json",
                         linker=EmbeddingLinker(embed=make_st_embedder()))

planner = Planner(store=store, kg=kg)

Better entity linking

from hubmesh.kg import build_entity_kg
from hubmesh.entity_linker import EmbeddingLinker, make_st_embedder

# Cluster surface variations: "United States" / "U.S." / "USA" → one entity
linker = EmbeddingLinker(embed=make_st_embedder(), threshold=0.82)
kg = build_entity_kg(docs, linker=linker)

Iterative multi-hop: let your agent drive

r1 = planner.retrieve(query=question, top_k=5)

# your agent reads r1, spots the bridge entity, then aims hop 2 at it:
r2 = planner.retrieve(
    query=question, top_k=5,
    seed_entities=["Nimbus Analytics"],           # merged with the query's own seeds
    exclude_docs=[s.doc.id for s in r1.sources],  # don't re-retrieve consumed docs
)

Seed mentions resolve through the alias index, so free-text entity names work. The query path stays deterministic and LLM-free — the planning intelligence lives in the caller.

MCP server: plug hubmesh into any agent

pip install "hubmesh[mcp]"
python -m spacy download en_core_web_sm

{"mcpServers": {"hubmesh": {"command": "hubmesh-mcp"}}}

Exposes the planner as deterministic operator tools over stdio — index_corpus, retrieve (seed-steerable, as above), resolve_entities, entity_neighbors, path_between, get_document, graph_stats, list_corpora. Your agent is the solver: it decomposes the question, reads each hop, and aims the next one; the server answers in milliseconds with zero LLM calls. Corpora persist as plain JSON/NPZ under ~/.hubmesh/corpora.

The server warms up models and persisted corpora in the background at launch (~5-10s on first run), so tool calls stay fast from the start — relevant for strict-timeout connector clients (Perplexity, etc.).

For web-based connector clients, serve SSE natively — no gateway process needed:

hubmesh-mcp --transport sse --port 8000 --allow-tunnel
ngrok http 8000     # paste https://<your-url>/sse into the connector

Tunnel field notes (from a live Perplexity integration): ngrok works (free tier included); cloudflared quick tunnels buffer SSE bodies and hang tool calls; supergateway is unnecessary here and crashes on reconnect. --allow-tunnel accepts the tunnel's forwarded Host header — without it, proxied requests get 421 Misdirected Request.

Full field report — setup, error decoder, a 9/9 test battery run through Perplexity, and two findings about reasoning-model behaviour — in docs/perplexity.md.

Chunking long documents

from hubmesh import chunk_by_sentences, chunk_documents

chunks = chunk_documents(
    [{"id": "doc1", "text": long_text}, ...],
    strategy="sentences", target_tokens=200,
)
# Then embed chunks and index normally

Installation

pip install hubmesh                   # core
pip install "hubmesh[qdrant]"         # Qdrant adapter
pip install "hubmesh[chroma]"         # Chroma adapter
pip install "hubmesh[kg]"             # entity-linked KG (spaCy)
pip install "hubmesh[linker]"         # embedding-based entity linker
pip install "hubmesh[all]"            # everything
python -m spacy download en_core_web_sm   # required for KG mode

Design

query → first-pass ANN  → induced subgraph → multi-component scoring
                              ↓                        ↓
                       community anchoring → Personalized PageRank
                              ↓                        ↓
                              └─────→ ranking → budget-aware packing → context

Each layer is independently testable and replaceable. Adapters wrap your existing vector DB so you don't have to migrate.

Benchmarks

Headline: on multi-hop QA, hubmesh's KG mode beats both naive cosine retrieval and a HippoRAG-style PPR-only ablation that uses the same KG, at every hop depth.

Benchmark	Setting	recall@10 vs naive
HotpotQA dev, N=7405 (full)	KG mode	+5.90 pts
HotpotQA dev, N=500	KG mode	+5.0 pts
MuSiQue dev, N=300, 2-hop	KG mode	+6.0 pts
MuSiQue dev, N=300, 3-hop	KG mode	+3.2 pts
MuSiQue dev, N=300, 4-hop	KG mode	+5.0 pts

All rows measured with v0.4.0 defaults (alias-indexed seeds + NNSI-KG convergence; ablation JSONs committed in benchmarks/). Disclosed: convergence trades top-rank precision for depth recall — recall@2 is −0.75 pts vs naive on full dev (dips ≤0.5 at smaller n); if you retrieve with top_k=2, set use_convergence=False. Multi-seed queries cost ~1.5–1.8× (still zero LLM tokens, deterministic).

vs PPR-only ablation on the same KG: +29.8 pts on HotpotQA at N=500 (measured on v0.2.0) — the multi-component scoring is doing the work, not just "having a graph."

On the full N=7405 HotpotQA dev: hubmesh hits 75.2% supporting-fact recall@10 vs naive cosine's 69.3% (+4.21 pts at recall@5; recall@2 −0.75, disclosed above).

Latency: ~22 ms mean / 26 ms p95 per query on a 7K-node KG (after PPR matrix caching); ~3 s/query at the 66K-paragraph full-dev scale with v0.4 convergence on.

See BENCHMARKS.md for the full methodology, ablations, per-hop breakdown, and notes on what this proves and doesn't.

Reproduce:

python benchmarks/run_hotpotqa.py --n 500 --kg
python benchmarks/run_musique.py  --n 300 --kg
python benchmarks/profile_query.py        # latency profile

Status

Pre-alpha (v0.4.0). Core algorithms implemented and validated; adapters for in-memory, Qdrant, and Chroma; entity-linked KG with both spaCy NER and LLM-based extraction (both linker-aware); alias-indexed entity resolution; NNSI-KG scoring (multi-source convergence default-on, hub-discounted PPR opt-in); agent-driven iterative multi-hop via seed_entities / exclude_docs; MCP operator server (hubmesh-mcp, native SSE) with JSON/NPZ corpus persistence; document chunking; reasoning-path explanation; PPR-cache latency optimisation. Pinecone / pgvector / Weaviate adapters and additional multi-hop benchmarks are tracked as good first issues.

Acknowledgements

The multi-component scoring pattern is adapted from the Network Node Significance Index (NNSI) framework introduced in Naidu Dsk, "A Framework for Improving Network Topology Based on Graph Theory in Software-Defined Networking", 26th International Conference on Internet Computing & IoT (ICOMP'25), Las Vegas, July 2025 — proceedings to appear. Repurposed here from SDN topology optimization to retrieval planning.

License

MIT