KnowledgeCore Deployment Guide

Hardware sizing

Pick the tier that fits your corpus.

KnowledgeCore runs the AI models locally. Sizing depends on how many documents you're indexing and how many concurrent users will query. All tiers support air-gap deployment.

Minimum

Proof of concept · Up to 500 pages

CPUApple Silicon M2 Pro, Intel i9-12th gen, or AMD Ryzen 9 equivalent RAM32 GB unified/system memory Storage512 GB NVMe SSD GPUIntegrated or no discrete GPU (CPU inference — slower) OSmacOS 14+, Ubuntu 22.04 LTS Concurrent users1–3 First-token latency15–25 seconds

Recommended

★ Best for production · Up to 5,000 pages

CPUApple Silicon M3 Max or M4 Max, AMD EPYC 9004 with NVIDIA A10G RAM64 GB unified (Apple) or 64 GB + 24 GB VRAM (Linux) Storage1 TB NVMe SSD GPUFull Metal (Apple) or CUDA 12.x (Linux) OSmacOS 15+, Ubuntu 22.04 LTS Concurrent users4–12 First-token latency3–7 seconds

Enterprise

High-volume · 5,000–50,000 pages

CPUApple Silicon Max or Ultra (Mac Studio) or dual-socket AMD EPYC RAM128–192 GB Storage2–4 TB NVMe (RAID-1 recommended) GPUApple Silicon GPU (Metal) or 2× NVIDIA A100/H100 OSmacOS 15+, Ubuntu 22.04 LTS, RHEL 9 Concurrent users20–100 First-token latency<3 seconds

Storage formula	Estimate
Per document page (PDF/DOCX)	~50 KB of vector storage in Qdrant
1,000-page corpus	~50 MB vector store + ~200 MB raw documents
10,000-page corpus	~500 MB vector store + ~2 GB raw documents
AI models (recommended tier)	~18 GB LLM (qwen3:30b-a3b) + ~1 GB embeddings (bge-m3) + ~8 GB auxiliary models
OS + Docker + application	~30 GB overhead
Recommended minimum free space	Documents + models + 2× headroom

Prerequisites

What must be installed before we arrive.

TechnoNet handles the KnowledgeCore installation. The following must be ready on the target machine before the setup session.

Requirement	Version	Notes
Docker Desktop (macOS) or Docker Engine (Linux)	27.x or later	Required All application services run in Docker containers
Docker Compose	v2.x (included with Docker Desktop)	Required
Available disk space	As per sizing table above	Required
Outbound port 443	—	Required For initial model download during setup. Can be blocked post-install for air-gap operation.
Python 3.11	3.11.x arm64/x86_64	Required For ingest pipeline (runs outside Docker)
macOS login to account with admin rights	macOS 14+	Required For launchd service registration
SMTP credentials (optional)	—	Recommended Enables licence email delivery and renewal reminders
Domain / hostname	—	Recommended For HTTPS access on your network. We can configure with an IP address if no domain is available.

Installation

What happens on setup day.

A TechnoNet engineer (on-site or remote via SSH/Screen Sharing) completes these steps. Estimated time: 4–6 hours for a first deployment including model downloads.

1

Deploy application containers

Docker Compose brings up 5 services: KnowledgeCore API, Qdrant vector database, PostgreSQL, Redis, Neo4j. Services are configured to restart automatically on reboot.

# Clone or copy the deployment package
git clone <repository URL supplied with your licence> /opt/knowledgecore
cp .env.example .env  # edit with your credentials
docker compose up -d  # starts all 5 services

2

Install and start the LLM inference server

llama-server (llama.cpp build) is registered as a system service via launchd (macOS) or systemd (Linux). It loads the AI model into memory and exposes the OpenAI-compatible API on port 11436. Model download: ~17 GB — takes 15–30 minutes on a fast connection.

# Model is copied to the models/gguf/ directory
# launchd plist is installed to ~/Library/LaunchAgents/
launchctl load ~/Library/LaunchAgents/com.knowledgecore.llama-server.plist
# Verify: model loaded, accepting connections
curl http://localhost:11436/health

3

Install embedding and proposition models

Ollama is installed as a login item (macOS) or systemd service (Linux). The bge-m3 embedding model (~600 MB) is pulled. The proposition extraction model (~4 GB MLX) is installed as a second launchd service.

ollama pull bge-m3:latest
ollama pull qwen3:14b   # on-demand — only loads on first call
# vllm-mlx for proposition extraction
launchctl load ~/Library/LaunchAgents/com.knowledgecore.vllm-mlx.plist

4

Configure tenant and ingest documents

A tenant configuration file is created for your organisation with document authority weightings. Your documents (PDF, DOCX, XLSX) are placed in the ingest directory. The ingest pipeline processes each document: PDF parsing → proposition extraction → embedding → vector storage → relational triple extraction.

# Copy your documents to data/raw/<tenant_id>/
# Run the ingest pipeline
python3 scripts/ingest.py --tenant your_org --all
# Monitor progress
tail -f tenants/your_org/outputs/ingest.log
# Typical time: 200-page corpus ~11 minutes

5

Verify end-to-end with a live query

A test query confirms the full pipeline: retrieval, reranking, and answer generation. Latency and source citations are checked. The health endpoint confirms all services are online.

curl -s http://localhost:8000/health | python3 -m json.tool
# Expect: "status": "ok", all services green

curl -X POST http://localhost:8000/query \
  -H "Authorization: Bearer YOUR_API_KEY" \
  -d '{"question": "What is our data retention policy?", "tenant_id": "your_org"}'
# Expect: cited answer with source documents listed

6

Configure network access and backup

Caddy reverse proxy is configured for HTTPS access. API keys are issued for end users. NAS backup is scheduled (nightly at 3 AM) — backs up PostgreSQL, Qdrant vector snapshots, and raw documents.

Deployment timeline

From sign-off to live system.

Day 0Licence issued, deployment package and credentials shared securely. Hardware prerequisites confirmed.

Day 1Setup session (on-site or remote, 4–6 hours). Containers deployed, models loaded, documents ingested, end-to-end verified. System is live.

Day 2–5Users begin querying. Any retrieval gaps or document issues surfaced and resolved. Additional documents ingested if needed.

Week 2Optional: RAGAS evaluation run (requires a declared maintenance window). Accuracy baseline documented for your corpus.

OngoingNew documents: re-ingest at any time. Model upgrades governed — benchmark before applying. Nightly backup verified via status file.

Maintenance

Keeping the system healthy.

KnowledgeCore is designed for unattended operation. The following procedures cover the most common maintenance tasks.

Task	When	Command / procedure
Health check	Any time	`~/hc.sh` — 29-check colour-coded report: API, LLM, vector DB, cache, NAS backup, all services.
Add new documents	When corpus changes	Copy files to `data/raw/<tenant>/`, then `python3 scripts/ingest.py --tenant <id> --new-only`. Flush Redis cache after.
API restart (after config change)	Config changes only	`docker compose up -d --force-recreate knowledgecore`
LLM server restart	If model hangs or swap file grows	`launchctl kickstart -k gui/$(id -u)/com.knowledgecore.llama-server`
Docker daemon freeze	Rare — symptom: all docker commands hang	Hard-kill all Docker processes, reopen Docker Desktop from GUI. Requires Screen Sharing if headless.
Verify nightly backup	Daily (automated via AuditCore monitor)	`cat nas_backup_status.json` — shows last backup phase and status.
Model upgrade	Governed — TechnoNet-approved only	Upgrade → RAGAS benchmark (maintenance window) → ≥95% pass → update compatibility matrix → deploy.
Application update	On new release	`git pull && docker compose build knowledgecore && docker compose up -d --force-recreate knowledgecore`

Document support

What you can ingest.

Format	Support	Notes
PDF (text-layer)	Full	Scanned PDFs without a text layer require OCR pre-processing (not included — contact us for the OCR add-on).
DOCX (Word)	Full	Headings, tables, and bullet structure preserved.
XLSX (Excel)	Full	Tabular data extracted row-by-row. Formulae values (not formulas) are indexed.
TXT / Markdown	Full	Plain text and Markdown are ingested natively.
Scanned PDF (image-only)	OCR add-on	Available as a professional services add-on. Adds Tesseract OCR pre-processing stage.
Images (charts, diagrams)	Roadmap	Multimodal ingestion (visual + semantic) is on the engineering roadmap. Currently only text content is indexed.
Audio / video	Not supported	Transcript pre-processing required before ingest. Contact us for a transcription pipeline quote.

Up and runningin one day.