Mium Key Features
Without Ontul, Mium is meaningless. Every capability — natural-language analytics grounded in certified definitions, job lifecycle, code generation — is built around Ontul as the data engine.
Natural-Language Data Analytics
Ask in plain language; agents answer with a table, a chart or a single headline figure, and export to CSV, Markdown, Excel, PDF or PPTX. The answer comes first and the SQL sits one click away, editable and runnable for the analyst who wants to take over.
Grounded in Certified Definitions
Mium keeps no semantic layer of its own. It aims the model at Ontul's certified metrics, dimensions, join paths and ontology, so the model picks names while the engine writes the aggregate and the JOIN — the difference between an answer that is correct and one that is merely plausible. Every answer shows which definitions it used and whether they are certified, resolved from the SQL that actually ran, so an answer written against raw tables says so.
Search SQL Cannot Express
"Waterproof products", "things like this one", "bought together, two hops out", "products that look like this photo" — questions about relevance and relationships rather than aggregation. Mium calls Ontul's certified retrievers (full-text, vector similarity, graph traversal, hybrid) and NeorunBase runs them as BM25, HNSW and graph traversal. Paste a screenshot and the worker's CLIP turns it into a 512-dimension vector that searches the same space as the text. Hits are shown as ranked results with scores and images, not as a grid, and whether re-ranking actually ran is reported honestly.
Answers That Improve With Use
When an answer is wrong, one click reports it; an analyst corrects the SQL and verifies it, and everyone asking that question afterwards gets the verified answer with the verifier's name on it. House rules written in plain language — the fiscal calendar, which rows to exclude — apply to every question without touching a metric definition. A benchmark suite of questions with known-correct answers says whether any of it actually helped.
Multi-Agent Orchestration
A strict-JSON action protocol dispatches fifteen actions in a single agent loop — answering from what the conversation already established, ad-hoc queries, batch and streaming job submission, status, logs and kill, job history, code generation, catalog management, and Ontul admin passthrough. Provider-agnostic; swap LLMs without changing application logic.
Pluggable LLM Backend
Connect the LLM you already use — Anthropic Claude or self-hosted Ollama. Credentials live envelope-encrypted in the ConnectionStore, and each user can pick a different model.
Ontul Job Lifecycle
Submit, monitor, log-stream, and kill long-running Ontul Batch and Streaming jobs from a chat conversation. Generate Java (Batch / Streaming / Class) and Python SDK source for jobs you want to ship into a pipeline.
Sovereign Operating Model
IAM, KMS, and ConnectionStore live in embedded RocksDB. Chat memory, prompts, embeddings, verified answers, house rules and benchmark history live in NeorunBase. Server-rendered files live in S3-compatible storage. Masters and Workers stay stateless for application data.
IAM & Envelope Encryption
AWS-style JSON policies over users, groups, companies and organizations; access keys plus short-lived STS credentials. AES-256-GCM envelope encryption protects every sensitive payload, with a versioned KMS key set you can rotate at runtime.
High-Availability Cluster
Master-Worker topology with ZooKeeper leader election and a custom NIO control plane. Followers transparently proxy writes to the leader and self-heal via snapshot pulls; Workers parallelize LLM calls and tool execution. Netty serves the Admin UI and REST API.
Integrated Admin Console
A React-based Admin UI unifies chat, IAM and policy editing, KMS key management, node topology and metrics, real-time log tailing, and S3 export storage configuration — all in one place.
Maintenance Mode — Close Writes, Keep Reads
For the moments when a write landing mid-operation is the problem — restoring a backup, rotating keys — the cluster refuses writes while staying readable. Agent turns, chat-session edits and job submission come back 503 with Retry-After; reads keep being served and IAM, KMS and connection settings stay editable. The agent turn heads that list because it is the heaviest write Mium makes: it persists chat memory and, through tools, reaches Ontul, GitHub, Slack and whatever else is connected — well outside Mium. The retention sweep pauses too. It is the one background loop that deletes data by TTL, and during a restore it would remove exactly what the restore had just brought back, indistinguishable afterwards from the restore having failed. Running jobs are not killed and the path to stop them stays open. If the setting row cannot be read the cluster is treated as open, not closed — a momentary database blip becoming a cluster-wide outage is the far worse failure.
Use Cases
Business Data Analytics
Business users without SQL skills query and analyze Ontul data in natural language — answering from certified metric definitions, with the definitions used shown beside every answer.
Ontul Job Operations
Data engineers submit, monitor and kill Batch and Streaming jobs from chat, and generate SDK source to land jobs in a pipeline faster.
Reports & Exports
Hand off analysis results as Chart, Markdown, Excel, PDF or PPTX — shareable reports on demand. Server-rendered files are envelope-encrypted at rest before download.
On-Premises Sovereign AI
Run LLM-driven analytics on your own infrastructure — data never leaves your network, and operations align with the rest of the CCL stack.
Considering Mium for your data platform?
Ontul-First · Sovereign AI · On-Prem.
Start natural-language analytics and job orchestration with an AI agent platform purpose-built for Ontul.