PAGE Architecture
DOC Enterprise Platform Reference
VERSION BGE-M3 Baseline
Step 1 — You upload
Drop in your documents
Upload your PDF, Word, and Excel files through the admin panel or API. KnowledgeCore reads every page, extracts all content, and builds a searchable knowledge base. Supported: PDPA policies, HR handbooks, compliance manuals, SOPs, technical specifications, financial reports. Processing happens on your server — files never leave your network.
Step 2 — Staff ask
Plain-language questions
Staff type a question in plain English — "What are our obligations if a vendor has a data breach?" — and receive a direct answer within seconds. No training required. No search keywords. No reading through 50-page documents. The system finds the right passage and drafts a clear answer from it.
Step 3 — Every answer cited
Traceable to the source
Every answer includes the exact document name, page number, and section it came from. Staff can click through to the original text. Legal and compliance teams can audit every answer. If the answer is not in your documents, KnowledgeCore says so — it will not invent information it has not read.
Step 4 — Nothing leaves
Complete data sovereignty
Unlike ChatGPT, Microsoft Copilot, or any cloud AI service — KnowledgeCore does not send your documents or queries to any external server. The AI model, the database, and all processing run on your own hardware. PDPA-compliant by design. Air-gap capable for the most sensitive environments.
Language support
English — production-ready: 100% on the published benchmark, 98% on a 43-question academic-policy benchmark. · Bahasa Malaysia — An initial Bahasa Malaysia evaluation (20 questions, June 2026) passed every question with no hallucinations, but its correctness score (0.585) is well below English (0.960); a full Bahasa Malaysia benchmark is being prepared. Contact us if your corpus is primarily BM or mixed-language.
Technical Architecture
The following section covers the full technical stack. Intended for IT architects, system integrators, and technical evaluators.
Layer 01
Memory
Tenant Configuration
Relational Facts (PostgreSQL)
Force-Inject Overrides
Authority Map
Conversation History
Entity Synthesis Pages
Long-Term User Memory →
Persistent knowledge context — what the system knows before a query arrives. Tenant-scoped, document-authoritative, and incrementally enriched with every ingest.
Layer 02
Intelligence
Proposition Extraction (vLLM-MLX)
BGE-M3 Embeddings (1024-dim)
3-Way Hybrid Search (Dense + SPLADE + BM25)
BGE Reranker v2-M3
Relational Lookup (24 relations)
Query Decomposition
Contradiction Removal
Adaptive Top-K
LangGraph Orchestration
AG-1 Retrieval Grader
AG-2 Answer Grader
AG-3 Coverage Guardrail
AG-4 Clarification Agent
AG-5 ReAct Tool Use
qwen3:30b-a3b Generation
Ten-node agentic pipeline — clarification, relational reasoning, retrieval, chunk grading, reranking, coverage check, inference, generation, and answer grading. AG-1 through AG-5 form a self-correcting accuracy loop. Every node is a discrete, replaceable skill.
Layer 03
Control
Pre-Ingest Validation
Licence Enforcement (JWT)
Authority Scoring
API Authentication (Bearer)
Pinned Dependency Versions
RAGAS Regression Gate
Automated Test Suite (pytest)
Model ID Validator
Guardrails at every boundary — ingest, query, and upgrade. No document enters without validation. No component upgrades without a signed-off benchmark pass.
Layer 04
Delegation
RAGAS Diagnostic Agent
Relational Fact Agent
Autonomous RAG Optimiser
GraphRAG Financial Traversal
Student Advisory Agent →
Specialised subagents handle tasks the core pipeline delegates — evaluation, knowledge extraction, configuration tuning, and domain-specific advisory workflows.
Layer 05
Distribution
REST API (FastAPI)
MCP Server
Docker Compose Package
Guided Installer (install.py)
Cloudflare Tunnel
ERP / Teams Integration →
KnowledgeCore runs where your data lives. On-premise or network-isolated — with an MCP interface that embeds it into any compatible tool or workflow.
01
Memory
What the system knows before you ask
02
Intelligence
How it reasons across your documents
03
Control
Guardrails at every boundary
04
Delegation
Specialised agents for complex tasks
05
Distribution
Runs where your data lives
Data Sovereignty & Compliance
PDPA Compliance
Your data never leaves your server.
KnowledgeCore is fully on-premise. All queries, documents, and model inference run on your hardware. There is no cloud dependency, no external API call, and no telemetry. Compliant with Malaysia's Personal Data Protection Act 2010 and Brunei's data residency requirements out of the box.
Regulated Industries
Built for environments where cloud AI is not an option.
Manufacturing with proprietary process documents. Financial services with client-sensitive data. Higher education with confidential policy records. Government with classified operational procedures. KnowledgeCore runs air-gapped if required — no internet connection needed after deployment.
No Vendor Lock-In
You own the stack. Cancel and keep everything.
Your documents, your models, your database. Nothing is held in TechnoNet infrastructure. Cancel at any time — your deployment continues to run. The codebase is delivered to you; the intelligence stays where you put it.
Accuracy Before Delivery
Benchmarked on your documents, not ours.
Enterprise deployments are validated using RAGAS — an industry-standard AI evaluation framework — run against real questions drawn from your actual corpus before handover. You approve the accuracy score before we hand over the keys.
Why KnowledgeCore
01
Your data never leaves your server.
Fully on-premise. No cloud dependency, no third-party data access. Full compliance with data sovereignty requirements — critical for regulated environments in Malaysia and Brunei.
02
Knowledge that compounds.
Every document ingested enriches the relational graph. Every query draws on structured reasoning — not just keyword retrieval. The system gets more capable with every upload.
03
Accuracy benchmarked before every change.
RAGAS evaluation runs against 43 hand-crafted questions before any component upgrade deploys. Production environments never change without a signed-off benchmark pass.
How Deployment Works
Step 01 · Install
One command. Nine steps.
python3 install.py
The guided installer walks through environment validation, dependency checks, database setup, model configuration, and licence activation. Runs on Linux (CUDA or CPU) and macOS Apple Silicon. Most installations complete in under 30 minutes.
Step 02 · Configure
Point at your documents.
POST /ingest
Upload your documents via the REST API or drop them in the ingest folder. KnowledgeCore extracts propositions, builds the relational graph, and indexes everything into the hybrid search engine. Supports PDF, DOCX, XLSX, and plain text.
Step 03 · Live
Query in minutes.
POST /query
REST API endpoint is live immediately after ingest. MCP server connects to Claude Desktop and compatible tools. The system answers questions with cited sources, relational reasoning, and full data sovereignty — entirely on your hardware.
BGE-M3 Production Baseline — 25 May 2026
Current Pass Rate
98%
42 / 43 questions passing (demo corpus)
15 / 15 passing (PDPA compliance corpus, 100%)
Quality Metrics
Faithfulness 0.977
Relevancy 0.737 · Correctness 0.938
bge-m3:latest (1024-dim) + llama3.3:70b via llama-server (the generator before the 27 May 2026 switch to qwen3:30b-a3b)
Deployment
On-Premise · Linux / macOS
Docker Compose · CUDA + CPU fallback
One-command installer · JWT licence-enforced
See it running on
your documents.
30 minutes. Bring your hardest document question. We will answer it live.