---

kompend

Knowledge-Oriented Markdown Processing and ENtity Distillation

A compiled LLM-wiki system for scientific and technical documents. Parse, chunk, resolve, and compile PDFs, DOCX, spreadsheets, and web sources into an interlinked, queryable knowledge base.

Unlike most tools in this space, kompend is provider-agnostic, adapts to whatever LLM you have — from a local 8B model to a frontier API — and uses structure-preserving parsing so technical content (tables, equations, figures) survives extraction intact. Read why →

---

5-Tier Pipeline

kompend follows a parser-agnostic, structure-preserving pipeline:

graph LR
  T1[Preprocessing] --> T2[Chunking]
  T2 --> T3[Resolution]
  T3 --> T4[Compilation]
  T4 --> T5[Output]

Quick Start

# Initialize a project
kompend init ./my-wiki && cd ./my-wiki

# Compile documents into a wiki
kompend compile *.pdf

# Query the compiled wiki
kompend query "what is the main result?"

Key Features

• Parser-agnostic — one configured backend per vault; parsecraft routes each page across its own sub-backends (native, LiteParse, docling, pandoc, OCR) and claims 32 suffixes. See full breakdown
• Hybrid chunking — Atomic elements + recursive token-aware splitting
• VLM enrichment — Image/figure captioning via vision-language models (--vlm-model)
• Typed knowledge graph — Entity relations with confidence (uses, depends_on, contradicts, …)
• Memory lifecycle — Confidence decay, tier promotion (working→episodic→semantic→procedural), supersession
• Hybrid search — FTS + vector + token-overlap fused with reciprocal rank fusion (RRF)
• Distill — Synthesize captured notes into coherent wiki pages
• Event hooks — Pluggable callbacks at compile/capture/query/lint lifecycle points
• WiCER compilation — Iterative compile-evaluate-refine loop for quality
• Multi-format output — Markdown (default, OKF v0.2-compliant), JSON, Typst
• LLM-powered linting — 8 structural rules + 4 LLM rules (incl. contradiction detection)

---

Documentation

Getting Started

• Motivation — Why kompend exists and what makes it different
• Getting Started — CLI setup, compile/query examples
• Input Pipeline — capture/learn/distill/consolidate: turning notes into wiki sources
• Initialization — kompend init project scaffolding
• Configuration — 5-layer config system, kompend config subcommands
• Basic Principles — What kompend does and why
• Document Extraction — Supported formats, parser selection, framework decisions

Pipeline & Architecture

• Architecture — 5-tier pipeline design
• Pipeline Tiers — Per-tier component reference
• Decision Log — Architectural decisions and rationale
• Caching — Multi-level content-addressed caching
• Cache Layout — .kompend/ directory layout
• Output Architecture — Output writers, renderer, flavor system

Compilation & Output

• Compilation Guide — Compilers, VLM, relations, memory, hooks
• Output Configuration — Markdown flavors, frontmatter, formatting
• Attachments — Supplementary files per source, opt-in ingestion
• WiCER Loop — Iterative compile-evaluate-refine quality loop
• Review Queue — Human-in-the-loop approval for wiki pages
• Wiki Structure — index.md, log.md, overview.md

LLM Configuration

• LLM Configuration — Provider setup, role routing, failover chains
• PydanticAI Agents & Skills — Agent architecture, per-role models, skill configuration
• Model Context Adaptation — Context window discovery, adaptive sizing

Linting & Quality

• Linting Guide — Structural + LLM-powered rules
• Provenance — Trace wiki pages back to their source files
• Trust & Memory — Provenance, memory lifecycle, decay
• Dedup — Content-level duplicate detection and suppression

Memory & Knowledge

• Memory Lifecycle — Confidence, tiers, decay, supersession
• Query Memory — Episodic memory, context synthesis, and MCP integration
• Distill — Captured notes → synthesized wiki pages

Integrations

• MCP Server — Agent integration via MCP
• ACP Agent — Editor integration via Agent Client Protocol
• Open-WebUI — Connecting kompend with Open-WebUI
• Ecosystem — LLM-wiki tool survey

Project & Community

• Architecture Decisions — Design choices and rationale
• Acknowledgements — Credits and investigated projects