Help & documentation

Everything,
indexed.

Find the right doc for your question — grouped by what you're doing, not by what the system is called. The pointers go to the deeper documents that already exist (API reference, features catalogue, sector pages, troubleshooting), not to duplicates.

Installation & first run

Standing up the platform.

Get started

Install in five minutes — Linux server, macOS workstation, or multi-node cluster. Three platform paths covered.

First run

From `dnf install` complete to a working cluster — admin signup, license, first KB, first chat.

Edge install

Single-node install on Raspberry Pi, Intel NUC, or NVIDIA Jetson. Offline operation with store-and-forward to a central cluster.

Install troubleshooting

Nine common failure modes (GPG fail, port conflicts, admin conflicts, Gatekeeper, license signature) with recoveries.

Cluster operations

Running the platform day-to-day.

Rolling upgrades (§70)

Cluster-wide RPM upgrades orchestrated through the controller. Drain → install → restart per node. Reference: api-reference.md §19.

Backup & DR (§40)

Local-destination snapshots of cluster state — controller, vector, memory, configs. Verify + restore per snapshot. Reference: api-reference.md §18.

PKI & ACME (§39)

Internal CA plus Let's-Encrypt issuance + renewal via certbot. Generate, deploy, rotate. Cluster-wide push from master. Reference: features §13.4.

4.x → 5.0 migration (§85)

Walks a 4.x data tree (vectors / memory / agent / comm / science) and replays into 5.0 syscall surface. Companion CLI: scripts/migrate-4x-to-5.0.sh.

Replication (§29)

Real rsync-over-SSH replication between data workers. Vector storage, matrix memory, file storage tenants. SSH key auth.

Webhooks (§75)

Outbound event publisher with HMAC-SHA256 signing. Failed deliveries auto-disable after threshold. Reference: for-developers.

OTLP export (§90)

OpenTelemetry exporter — spans, counters, histograms. Off by default; admin opt-in. Low-cardinality path normalisation built in.

Audit ledger

Hash-chained log of every AI-assisted decision. Defensible record for compliance reviews. Reference: features §13.5.

Cluster dashboards

Controller serves controller-overview, swarm-management, cluster-ops, and knowledge-browser dashboards. Access at /dashboard.

Chat shell

Using the chat.

Model providers

Every model in the picker shows a small coloured badge identifying the backend that serves it — Ollama, OpenAI, Eldric Inferenced, vLLM, llama.cpp, and the rest. Green for cluster-local, brand-coloured for external. The same badge appears on every assistant message.

Show tool query details

Chat settings → Inference → toggle Show tool query details. When on, every tool call shows a one-line summary of what was queried (knowledge base, science source, web search, etc.) above the result. Default on; turn off for a denser transcript.

Smart memory inference

Opt-in per request via smart_memory: true or per-tenant admin toggle. The model consults your associative memory at the prompt boundary; answers stay grounded in your own data without a separate retrieval call.

Themes follow the user

Per-user theme settings sync across all clients (Mac GUI, iOS, iPad, web) via the controller. Settings follow the user, not the device.

Notifications

Long-running agent tasks surface as in-app notifications with progress and one-tap return. Apple Push wires through to lock-screen alerts on iOS and iPad.

Eldric on iPad

Floating composer, Apple Pencil + Scribble, drag-drop ingest from Files / Photos / Mail, multi-window via Stage Manager. TestFlight today; App Store next.

Developing

Building against Eldric.

For developers (overview)

OpenAI-compatible API, plugin authoring, webhooks, agents — drop-in replacement for OpenAI tooling.

api-endpoints

Complete endpoint surface — Edge, Controller, Router, Data, Agent, Media, Comm, Science, Training, IoT, Swarm, NOVA, xLSTM, Cluster Operations.

api-public

Edge-reachable endpoints only — one line per endpoint. The starting point for external integration.

Plugin authoring (§80)

Five plugin types: Tool, Filter, Pipe, Action, Widget. Python or JS. Manifest + valves + main code. Marketplace at /api/v1/marketplace/catalog.

Agent Builder + Generator

Agent Builder uses LLM reasoning to design novel agents; Agent Generator uses domain templates for deterministic mass production. Both write to agents/<name>/.

Webhook subscribers

Subscribe at /api/v1/webhooks/subscriptions. HMAC-SHA256 signing on every outbound POST. Failed-delivery auto-disable.

Working with data

RAG, training, knowledge.

Using RAG

The customer-facing how-to — upload documents, ask grounded questions, read citations, manage knowledge bases. RAG is on by default in 5.0.

RAG architecture

Under the hood — Controller → Inferenced (embeddings via GGUF) → Data Worker (vector store). Configurable via ELDRIC_EMBED_BACKEND_URL.

Chunking strategies

Content-aware chunking with intelligent upload — different defaults per content type (PDF, code, CSV, audio, video, sensor). Suggestion → preview → commit.

RAG on demand

The four-tier cascade (ENRN → EMM → RAG → live) and the retention loop (search → accept → ingest → enrich → dream → train). Smarter with use.

Custom classification

Teach the router your own intent classes. Path A overlay classifier (with labelled examples); Path B LLM fallback (with your taxonomy). Pro+.

Matrix memory

Hierarchical associative memory — domain → project → run. Associative updates + single-step recall. Portable .emm format. Reference: api-reference.md §7.

Training pipeline

Six backends (Unsloth, Axolotl, TRL, DeepSpeed, MLX, llama.cpp), eight methods (LoRA, QLoRA, SFT, DPO, RLHF, PPO, full fine-tune, distillation). Reference: api-reference.md §13.

Distillation (§50)

Model → EMM knowledge distillation. Vector source → LLM Q&A pairs → matrix memory associations. Sync run via /api/v1/distill/run.

Federated learning (§44)

Round orchestration — controller broadcasts to participants, workers train locally, shards aggregate. Reference: features §9.5.

By worker type

Reaching deeper into a specific component.

Edge gateway

TLS termination, auth, rate limits, embedded webchat, plugin host. Port 443/80. Reference: api-reference.md §1-3.

Controller

Cluster orchestration, licensing, topology push, cluster operations. Port 8880. Reference: api-reference.md §18-19.

Router

Intelligent request routing — eight load-balancing strategies, intent classification, theme detection, AI-powered decisions. Port 8881. Reference: api-reference.md §5.

Worker (inference)

Multiple backends: Ollama, vLLM, TGI, Triton, OpenAI-compatible, llama.cpp. Port 8890. Reference: api-reference.md §4.

Cloud Worker

Multi-backend cloud gateway — Ollama Cloud, OpenAI, Anthropic, xAI, Groq, DeepSpeed, more. Port 8889.

Data Worker

File storage, NFS, databases, vector/RAG, matrix memory. Multi-tenant with quotas. Port 8892.

Agent Worker

Agentic RAG, multi-agent orchestration, query decomposition, training data generation. Port 8893.

Media Worker

STT (Whisper, OpenAI), TTS (Piper, ElevenLabs, OpenAI), audio analysis, video processing, voice chat. Port 8894.

Communication Worker

Email, SMS, WhatsApp, Signal, Teams, XMPP, VoIP. AI auto-response. Port 8895.

Science Worker

Source Registry (§43) — 16 categories, 28 built-in sources. Bioinformatics, pharma, CRISPR, LIMS. Port 8897.

Training Worker

Six backends, eight methods, training chains, latent reasoning. GPU inventory + multi-GPU. Port 8898.

Inference Worker (native)

Direct GGUF / xLSTM loading without Ollama / vLLM. Speculative decoding, continuous batching, pipeline parallelism. Port 8883.

IoT Worker

Consumer IoT (Netatmo, HomeKit, Matter) + industrial protocols (OPC-UA, Modbus, MQTT Sparkplug B). Port 8891.

Swarm

Multi-agent orchestration with six topology types (hierarchical, P2P, ring, star, mesh, hybrid). Port 8885-8887.

NOVA (experimental)

Goal system, tri-memory, reasoning, meta-learner, sandboxed self-modification. Port 8899. Off by default.

Compliance & security

What we explicitly support.

Multi-tenant isolation

Per-tenant boundary on storage, KBs, agents, projects. Cross-tenant access denied at the gateway. Tenant guard documented in features §16.1.

RBAC

Four roles — Viewer, Developer, Admin, SuperAdmin. Capability-gated API surface. Per-tenant assignment.

Audit ledger

Hash-chained log of every AI-assisted decision. Defensible record for HIPAA / GDPR / 21 CFR Part 11 reviews.

Compliance modes

GLP and FDA 21 CFR Part 11 compliance flags on the Science worker. IRB-conformant audit trail.

Air-gap deployment

Full on-premises stack with no external dependencies. Cloud workers and federated learning are optional.

Licensing

Ed25519-signed license files. Free, Standard, Professional, Enterprise tiers. Hardware-binding optional. License email: license@core.at.

By sector

Sixteen industry pages.

Per-sector workload patterns, compliance support, deployment shape. Each page has a workloads grid, a compliance row, and an honest-scope statement.

Health & life sciences: healthcare, pharma, genomics, CRISPR, neuroscience
Finance & legal: banking, insurance, trading, legal
Industry & mobility: smart manufacturing, automotive, robotics, logistics
Science & research: space, climate, materials
Troubleshooting & support

When something goes wrong.

Reference catalogues

The two big documents.