KnowLP-RAG — 7-Layer Retrieval Pipeline

Dual knowledge graph (P/S-Agent) + 4-engine unified search + feedback loop

KnowLP-RAG: 7-Layer Retrieval Architecture Layer 1 — Markdown Source Obsidian · Logseq · Joplin · plain .md folder Layer 2 — Metadata Extraction YAML frontmatter · tags · wiki-links · headings Layer 3 — Dual Graph Builder P-Agent: 555 prerequisite edges (read A before B) S-Agent: 624 similarity edges (viable alternatives) Layer 4 — Paragraph Chunking 542 chunks · body-text keyword matching · Chinese-aware sentence splitting Layer 5 — Vector Index (dual-mode) n-gram fast (~1s, CPU) · Qwen3-VL-Embedding-2B real embedding (~15s, GPU, 305×2048dim) Layer 6 — 4-Engine Unified Search KnowLP Graph P/S-Agent traversal Chroma Skills Vector skill search ripgrep Full-Text Raw content grep PixelRAG Visual search Layer 7 — Weight Feedback Loop consumed edges +0.05 (cap 2.0) · ignored -0.02 (floor 0.05) cold decay ×0.95 (30d unused) · graph gets better with every search feedback loop benchmarks 306 notes 1179 edges P@5: 0.407 MRR: 0.617 Legend Source / I/O Processing Graph / Index Search Vector Feedback Loopback KnowLP-RAG v3.0.0 · MIT · 36 tests · github.com/wly8691-jpg/knowlp-rag

Dual Graph (unique)

  • • P-Agent: prerequisite chains
  • • S-Agent: similar alternatives
  • • 1179 total edges, 306 nodes
  • • Jaccard + tag + dir co-occurrence

4-Engine Search

  • • KnowLP graph traversal
  • • Chroma skill vector search
  • • ripgrep raw content grep
  • • PixelRAG visual search

Feedback Loop

  • • +0.05 per consumed edge
  • • -0.02 per ignored edge
  • • ×0.95 cold decay (30d)
  • • Gets better with every use

KnowLP-RAG · github.com/wly8691-jpg/knowlp-rag · MIT © 2026