CSED Knowledge Base¶
The CSED Knowledge Base is a federated cognitive system that serves as the lab's living memory — a structured, versioned repository of architectural decisions, research context, design rationale, and operational knowledge.
Architecture¶
The system is built on three core components:
- OKF (Open Knowledge Format) — A Markdown + YAML frontmatter standard for representing knowledge nodes, validated against the Google Open Knowledge Format specification (released June 2026). Each node carries a dual-label
typearray for faceted classification. - kb.toml Manifests — Dependency manifests using
registry://URIs that declare static and dynamic knowledge dependencies, resolved via$KB_ROOTand managed with thekbCLI toolchain (Knowledge Base Manager / KBM). - Hybrid GraphRAG — A dual-store retrieval architecture combining Qdrant (vector embeddings for semantic similarity search) with Neo4j (graph traversals for relationship-aware queries). This enables both "find documents similar to this one" and "trace the chain of decisions that led to X" queries.
Design Principles¶
The KB architecture is grounded in the Interpretable Context Methodology (ICM) (arXiv:2603.16021v2), which formalizes the principle that folder structure is agent architecture.
The Input/Output Firewall: A strict separation between immutable ground-truth assets (sources/ or db/) and mutable AI/human iterations (work/ or workspace/) is mandatory. This prevents ingestion pipelines from feeding unverified AI-generated drafts back into the knowledge graph, avoiding recursive hallucination loops. Files in sources/ are indexed as authoritative nodes; files in workspace/ are tagged as derivative work with DERIVED_FROM edges pointing back to their sources.
The 3-Tier Knowledge OS: Knowledge is organized across three tiers — Project Level (active, 1-6 months), Domain Level (permanent discipline knowledge), and Master Research Vault (global synthesis). This prevents monolithic knowledge bases from becoming unwieldy and allows agents to scope their context appropriately.
Organizational Memory: AI chat logs and web clippings are treated as first-class citizens alongside formal papers. The KB schema includes AIChatLog and WebResource node types specifically to capture infrastructure commands, design debates, and cluster quirks that never make it into publications.
Entity-Relationship Schema¶
The knowledge graph defines seven primary node types — ResearchPaper, Concept, Project, Methodology, Dataset, AIChatLog, and WebResource — connected by standard predicates: addressesProblem, usesMethod, evaluatesOn, implements, supportsClaim, and cites. These directional verbs turn the isolated database into a traversable semantic graph, enabling queries like "find all concepts evaluated by methodology X in research paper Y."
Key Design Decisions¶
| Decision | Choice |
|---|---|
| Registry | ArangoDB (KnowledgeNodes / KnowledgeEdges) — no flat kb.json file |
| Addressing | registry:// URIs resolved via $KB_ROOT |
| Edge syntax | [predicate::Display Title](href) — href authoritative, bare links default to discusses |
| Sync semantics | kb sync respects kb.lock; kb update re-resolves dynamic deps |
| Versioning | Append-only nodes (version / supersedes / is_latest) ≅ prov:wasRevisionOf |
| Vault storage | Box (interim) → Ceph S3 (target) |
Implementation¶
The Knowledge Base is implemented through two repositories:
- csed-ucm/kb — The pilot repository containing architectural specs, OKF schemas, dual-label ontology definitions, glossary, and strategy documents. This is the canonical reference for the KB architecture.
- csed-ucm/kbm — The Knowledge Base Manager (KBM) — a Python CLI tool (
kb sync,kb update,kb init) that serves as the orchestration engine. It parseskb.tomlmanifests and OKF Markdown files, extracts typed nodes and directional edges, and manages synchronization with the graph registry (ArangoDB). Currently in active development.
Document Parsing Pipeline¶
The KB ingestion pipeline is informed by ongoing research into state-of-the-art document parsing. See the Document Parsing SOTA report for a survey of DeepSeek-OCR 2, Marker, DELM, and other 2026 frameworks.
Supplementary Reading¶
- From Retrieval to Resilience: Understanding the LLM Wiki Concept — The foundational concept behind persistent, compounding knowledge bases: treating your knowledge base as a codebase with an LLM as "programmer," solving the fragmentation problem of standard RAG.
- Operational Guide: Building a Compounding Knowledge Base with LLMs — A practical three-layer architecture guide (Immutable Raw Sources → LLM-Owned Wiki → Schema Configuration) with Ingest, Query, and Lint operations for a near-zero-maintenance compounding knowledge base.
- The Anatomy of an AI Agent: From Static Models to Active Partners — A primer on "The Loop," agent harness infrastructure, context management, and delegation — the architecture that powers agentic KB interactions.