Weknora Graph RAG & Hybrid Retrieval Engine

Weknora Graph RAG & Hybrid Retrieval Engine

@weknora/knowledge-base·v1.5.0·storage

An enterprise hybrid RAG engine combining dynamic Knowledge Graphs with dense vector retrieval. Delivers multi-hop reasoning, relationship mining, and chunk deduplication for zero-hallucination Q&A.

Graph-RAG · Hybrid-Search · VectorDB · KnowledgeBase

GitHubNPM
Installation & Integration
bash
manager:
$ pnpm add @weknora/knowledge-base

Architecture

Weknora provides DeepSeek Harness agents with an enterprise-grade factual foundation. Traditional vector similarity search often fragments logical relationships, faltering when queries demand multi-hop reasoning across interconnected documents. This plugin implements a dual-layer hybrid architecture: dense vector embedding for rapid semantic candidate filtering, coupled with graph topological traversal for relational inference. It turns messy internal wikis and API manuals into clean, queryable knowledge graphs with verbatim source attribution.

Core Features

01

Hybrid Graph + Vector Search: Unites dense embeddings with topological relation traversal

02

Automated Entity Disambiguation: Normalizes technical synonyms and aliases to prevent graph drift

03

Verifiable Line-Level Citations: Every synthesized claim anchors directly to source file line numbers

04

Incremental Live Updates: Rebuilds local subgraphs instantaneously without full re-indexing

Core Workflow

01

Heterogeneous Ingestion

Parses Markdown, PDF, and code repositories into semantic AST chunks.

02

Entity-Relation Extraction

Extracts named entities and dependency edges into a property graph database.

03

Hybrid Dual Retrieval

Executes dense vector search and multi-hop graph walks with RRF reranking.

04

Attributed Synthesis

Injects reranked evidence into context, generating verifiable answers with citations.

Configuration Parameters Reference (YAML / JSON)

ParameterTypeDefaultDescription
retrievalModestringhybridRetrieval strategy mode
maxHopCountnumber2Maximum graph traversal hops
topKnumber5Number of top reranked chunks injected