What Eldric 5.0
actually ships.

The distributed kernel.

Microkernel-style runtime

A C++ kernel hosts independent modules (edge, controller, router, data, agent, media, comm, science, training, inference, native inference, xLSTM, IoT, swarm, NOVA) — each on its own port.

Topology push

The controller pushes the current cluster topology on every heartbeat. Workers auto-discover swarm, data, peer, router, agent and media worker URLs.

Tenant guard

Header-only kernel hook returns 403 on cross-tenant attempts across data, storage, vector, memory, agent, comm, swarm and tenant paths.

Thirty-plus LLM backends.

Unified backend layer

Eleven backend types: Ollama, vLLM, TGI, llama.cpp, MLX, NVIDIA Triton, TensorFlow Serving, TorchServe, ONNX, OpenAI-compatible, Eldric Cluster pass-through.

Cloud aggregation

One endpoint federates OpenAI, Anthropic, xAI/Grok, Together, Groq, DeepSeek, Mistral, Cohere, Fireworks, Perplexity. Priority routing with fallback.

Native inference

Inferenced loads GGUF and xLSTM models directly via embedded llama.cpp — no Ollama, no vLLM. Multi-GPU tensor split, speculative decoding, continuous batching.

Eight strategies, AI-controlled when you ask.

Intent classification

Router classifies every request into 13 intents (Chat, RAG, AgentInvoke, Swarm, MemoryStore/Recall, Data, Science, Media, Comm, Training, IoT, Admin) and forwards to the right worker class.

Theme detection

Medicine, legal, code, finance, science, creative, general. Each theme can carry its own default model and per-rule overrides.

Ensemble

Fan out a request to multiple models, then synthesise the answers through a designated synthesiser model. Useful for high-stakes decisions.

Storage, vectors, matrix memory.

Multi-tenant storage

Per-tenant file storage with quotas. Chunked upload protocol with 4 MB chunks and 24 h TTL on incomplete uploads.

Vector / RAG

SQLite, FAISS, ChromaDB or in-memory backends. Hybrid BM25 + vector search. Auto-chunk on ingest. Re-embed on document edit.

Matrix memory

mLSTM-inspired outer-product memory. Compressed recall alongside the exact vector store. .emm v3 binary format with WAL + checkpoint.

Grounded answers, on by default.

Using RAG

RAG is on by default in 5.0. Upload PDFs, DOCX, code, CSVs, audio, video, sensor streams. Ask grounded questions. Read citation chips that point back to the source passages.

RAG architecture

Controller routes; the native inference daemon embeds with a GGUF model (~80 MB, runs on CPU); the data worker stores chunks alongside vectors. Three processes, three responsibilities, one wire.

Content-aware chunking

Twelve default strategies — semantic for scientific PDFs, function-boundary for code, per-row for CSVs, per-utterance for audio, per-window for sensor streams. The intelligent upload flow suggests parameters; the operator confirms.

RAG on demand

Four-tier cascade — ENRN learned weights → EMM associative memory → RAG → live external sources. The retention loop turns accepted answers into the next training corpus over time.

Custom classification

The router classifier ships with 128 built-in classes. Add your own intent classes — overlay-trained from labelled examples or LLM-fallback with your taxonomy. Pro+.

On-prem embedding

The embedding model runs locally on Inferenced (or any OpenAI-compatible /v1/embeddings endpoint configured via ELDRIC_EMBED_BACKEND_URL). Documents never leave the cluster.

Fifteen agent types, agentic RAG, workflows.

General, Researcher, Coder, Validator, Planner, Analyst, Explorer, Runner, Searcher, Database, Learner, Network, Spider, Email, Ansible.

Agentic RAG (ReAct)

Iterates Thought → Action → Observation up to a configurable cap. Tools include vector search, web fetch, file read, and any swarm-registered tool.

Orchestration patterns

Sequential, parallel, MapReduce, dependency-graph. The orchestrator picks the right pattern by the workflow shape.

Training-data generation

Walks a knowledge base and emits LoRA-ready JSONL — code_qa, chat, alpaca, dpo. Used to bootstrap router training and domain adapters.

Seven messaging protocols, full voice pipeline.

Seven protocols

Email (IMAP/SMTP), SMS (Twilio), WhatsApp (Business API), Signal (E2E), Microsoft Teams, XMPP, VoIP (SIP/RTP). One unified message envelope.

STT & TTS

Whisper.cpp, OpenAI Whisper, Faster-Whisper for transcription. Piper, ElevenLabs, OpenAI for synthesis. Full telephone-style AI calls are in development; today the platform handles dictation, meeting transcripts and accessibility inputs reliably.

Multimedia RAG

Audio + video content indexed and searchable. Used by the comm worker for voicemail recall and by the chat shell for inline media references.

Sixteen-category source registry.

Sixteen categories: open access papers, space, particle physics, genomics, neuroscience, medical, chemistry, earth, climate, astronomy, archaeology, legal, patents, funder, industry, custom.

Source registry

One entry per data source. Admin toggles sources; users see only the enabled ones. The custom category is the plugin entry point — no code changes required.

11 LLM tools

Five user tools, six admin tools. Filtered by role. List sources, request activation, dispatch a query, manage credentials, approve / reject pending requests.

Specialty pipelines

Bioinformatics (BLAST, variant calling), pharmaceutical (docking, ADMET, AlphaFold), CRISPR (guide RNA, off-targets), LIMS (GLP, 21 CFR Part 11).

Six backends, eight methods, federated.

Backends

Unsloth (CUDA, 2× LoRA), Axolotl (YAML), TRL (RLHF/DPO), DeepSpeed (multi-GPU), MLX (Apple Silicon), llama.cpp (GGUF). xLSTM training runs via the xLSTM daemon (below).

Methods

LoRA, QLoRA, SFT, DPO, RLHF, PPO, full fine-tune, distillation. Plus latent-reasoning techniques: COCONUT, Quiet-STaR, pause tokens, hidden CoT, DeepSeek DSA.

Federated learning

Multi-round federated training across worker nodes. Controller broadcasts cluster://training/federated/{job}/round-N; workers train locally; gradients aggregate without sharing data.

Structured ML, four workload classes.

Policy execution

Closed-loop control policies (LRAM) drive real-time control over WebSocket, Modbus, OPC-UA and MQTT-Sparkplug-B. Watchdog-driven safety fallback when the policy misses its deadline.

Forecast + encode

Time-series forecasting (TiRex) on telemetry windows. Vision-language encoding (ViL) for perception tasks. Both license-gated per workload, structured error responses on missing capability.

Associative retrieve

Native C++ Hopfield-style retrieve backend — microsecond latency on CPU alone. Used by the router for fast classification and by the data worker for fuzzy recall.

Details on xLSTM & IoT transports.

Consumer + industrial in one worker.

Consumer IoT

Netatmo (weather, security), HomeKit, Matter. Device pairing and attribute read/write over the IoT worker's API.

Industrial protocols

OPC-UA for PLCs, SCADA, DCS. Modbus TCP/RTU for legacy equipment. MQTT Sparkplug B. Alarm management, time-series historian, OEE analytics.

Predictive maintenance

Live tag values flow into matrix memory. An inference worker runs anomaly detection and emits maintenance scores.

Where Eldric runs, who talks to it.

Eldric for iPad

Universal-app shape — NavigationSplitView, floating composer, Apple Pencil + Scribble, drag-drop ingest, Stage Manager and Split View multi-window. TestFlight today, App Store next.

Mac, iOS, web, CLI

The native Mac GUI runs auto-update via Sparkle. iOS ships in the same universal app as iPad. The chat shell is embedded in the edge gateway at /chat — no external client needed.

Edge runtime

Single-node Eldric on Raspberry Pi 4, Intel NUC, NVIDIA Jetson. Local chat, local matrix memory, store-and-forward when the central cluster is unreachable. ARM64 minimal RPM.

.nexus bundle export / import

Pack the matrix memory, vector documents, knowledge bases, classifiers and identity overlays of one Eldric installation into a single signed file. Move between installations with a clean merge.

Cross-distro packaging

Signed RPMs for Fedora 42+, Fedora 40, RHEL 9+, Rocky 9, Alma 9, CentOS Stream 9+, ARM64. macOS PKG with auto-update. Native Ubuntu 24.04 + Debian 12 DEBs follow shortly after GA.

Model provider badges

Every model in the chat picker carries a coloured provider badge — Ollama, Inferenced, OpenAI, Anthropic, xAI, HuggingFace, Groq. Green for cluster-local, brand-coloured for external APIs.

What ops needs.

Rolling upgrade

Drain in-flight requests, snapshot state, install, verify SHA-256, restart, validate, move on. Master fans out to peers under cluster secret.

PKI & ACME

Internal CA plus Let's Encrypt via certbot. Generate, deploy, rotate. Cluster-wide push from the master.

Per-tenant theming

Each tenant has its own theme — colours, fonts, sidebar layout — plus optional logo. Public GET, admin PUT, server-side HTML sanitisation.

Backup & DR

Snapshot of controller state, vector storage, matrix memory, tenant configs, license, edge plugins. Idempotent restore.

OpenTelemetry

Opt-in OTLP-HTTP exporter for spans, counters, histograms. Low-cardinality path normalisation.

Webhooks

Outbound webhooks with HMAC-SHA256 request signing. Failed deliveries auto-disable after threshold.

Plugin marketplace

Browse, install (SHA-256 verification + manifest validation), uninstall, update. Served from the edge module.

Distillation

Model → EMM knowledge distillation. Source chunks become Q+A pairs by an LLM, both sides embedded, the pair written as an outer-product association into matrix memory.

Dream engine

Pulls completed sessions, extracts themes via an LLM, ingests them into matrix memory. Cadences: manual, hourly, nightly, continuous, on-idle.

Plus chunked upload, mDNS discovery, tenant guard, 4.x → 5.0 migration, and the artefact store. The release notes walks the formal list; the API reference documents every endpoint behind these features.