The Ultra-Compact Graph-Vector Database
for GraphRAG & Multi-Agent Swarms

Co-locating 2/4-bit quantized vector search, SQLite FTS5 keyword indexing, and C-level recursive graph traversal in a single, serverless, micro-memory Python package.

Storage Compression

14.5x

▲ vs. Raw PDF source document

Average Retrieval Latency

23.3ms

▲ Sub-25ms hybrid 2-hop search

Contextual Recall Gain

+12.5%

▲ Over isolated pure vector search

Concurrent Throughput

151 ops/s

▲ 100% thread-safe atomic transactions

🧠 Live Multi-Agent Knowledge Graph Explorer

Nodes: 15 Edges: 18 Latency: 1.8 ms Engine: SQLite Recursive CTE

STAGE 01

Zero-RAM BM25 Search

Native SQLite FTS5 retrieves exact keyword matches with zero Python memory overhead.

SQLite FTS5 C-Engine (~1.5ms)

STAGE 02

Recursive CTE Graph Expansion

Pushes multi-hop neighbor discovery ($k$-hops) into a single C-compiled SQL statement.

SQLite Recursive CTE (~0.8ms)

STAGE 03

Quantized Vector Reranking

TurboVec 4-bit/2-bit index calculates cosine similarities strictly on graph-expanded candidate nodes.

TurboQuant Bitmask (~2.1ms)

Why Q-VEX Outperforms Existing Architectures

Comparing traditional decoupled vector stores and heavy enterprise GraphRAG setups against Q-VEX's unified hybrid engine.

Architecture Dimension	Standard Vector DB (FAISS / Pinecone)	Heavy GraphRAG (Neo4j + LLM)	Q-VEX (Unified Engine)
Storage Footprint	High (Float32 uncompressed vectors)	Extreme (Separate DB + Graph server)	14.5x Smaller (2/4-bit TurboVec)
Relational Knowledge	Topology-Blind (Isolated points)	Full Graph (Requires server)	Recursive CTE (Single local DB)
Cloud API Ingestion Cost	$0.00 (Vector only)	$$$ (Thousands of LLM calls)	$0.00 (Zero-Shot GLiNER Local)
Multi-Hop Traversal Speed	N/A (Not supported)	150–600 ms (Network roundtrips)	<25 ms (In-process C execution)
Multi-Agent Concurrency	Read-heavy, complex sync	External lock manager needed	Atomic RLock (151 ops/s safe)

⚡ 3 Lines to Agentic Shared Memory

pip install qvex[langchain,llamaindex]

from qvex import QVEX from qvex.integrations.langgraph_adapter import QVEXSemanticMemorySaver # 1. Initialize high-performance 4-bit hybrid store qvex = QVEX(dim=384, storage_dir="./agent_shared_memory", bit_width=4) memory_saver = QVEXSemanticMemorySaver(qvex_instance=qvex, embed_fn=my_embedder) # 2. Agent A writes discovery with relational graph edge mem_id_1 = memory_saver.save_memory("Raft consensus ensures leader log replication.", metadata={"agent": "researcher"}) mem_id_2 = memory_saver.save_memory("Q-VEX achieves 151 ops/sec concurrent multi-agent throughput.", metadata={"agent": "researcher"}) qvex.add_edge(mem_id_1, mem_id_2, edge_type="technical_dependency") # 3. Agent B retrieves multi-hop expanded insights in 2.6 ms insights = memory_saver.retrieve_memory(query="consensus throughput", k=3, hops=2)

The Ultra-Compact Graph-Vector Databasefor GraphRAG & Multi-Agent Swarms