Core Concepts
The mental model. Ten minutes of reading that will save you hours of orientation. If you only read one docs page, read this one.
One platform, many products
StellarBase is a modular platform with three complementary products. You can use any one of them standalone; they multiply when used together.
| Product | What it is | When to use |
|---|---|---|
| StellarBase | Core platform — knowledge base, agents, workflows, chat, search, collaboration | You have data + want to build AI workflows on top of it |
| StellarCloud | Managed open-source inference — OCR, embeddings, NER, rerankers, LLMs | You need AI models but don’t want to host them |
| StellarGate | Privacy proxy — anonymizes any request before it reaches a third-party LLM | You want GPT-4 / Claude / Gemini but can’t send real data |
Three building blocks inside StellarBase
1. Knowledge base
The knowledge base is your unified data layer. Every document, every email, every image, every row of a database you connect — all ingested, parsed, indexed, and linked. It’s not a storage layer; it’s a semantic layer that sits on top of your existing systems.
Key properties:
- 50+ connectors — Google Drive, SharePoint, iManage, Slack, Teams, Notion, databases, S3, plus custom connectors
- Every format — PDF, DOCX, images, audio, video, code, logs, email — parsed automatically by StellarOCR
- 24+ languages — content in any European language, searchable in any other
- Continuous sync — as sources change, the knowledge base follows
2. Agents
An agent is a scoped AI specialist you configure. You write its system prompt (this is where your domain knowledge lives), you choose its tools, you scope its knowledge. Every answer is cited to source.
Agents are the main way you encode your company’s playbook into the platform. A law firm builds a “MSA Reviewer” with the firm’s red-lines. A hospital builds a “Guideline Agent” that knows NCCN inside out. A manufacturer builds a “Rotating Equipment” agent that mirrors the senior foreman’s reasoning.
3. Workflows
A workflow is a deterministic pipeline. You arrange nodes (data sources, models, agents, outputs) in a graph. Same inputs produce the same outputs, every time. Workflows can be visual (drag-and-drop) or declarative (described to an agent in plain language).
Workflows are how you automate. An insurance broker runs “portfolio review” weekly. A government ministry runs “beneficiary oversight” nightly. A manufacturer runs “predictive maintenance” every 15 minutes. The workflow is the same; it executes against fresh data.
Key mental shifts
It’s a platform, not a specialized tool
StellarBase does not come pre-configured for “legal review” or “medical literature” or “contract drafting”. It comes with the building blocks you use to assemble those. This is deliberate — your workflow is different from your neighbour’s, and a platform that bends to both is more valuable than two tools that each cover half.
The practical consequence: the first week you use StellarBase is about configuration, not training. You connect sources, write agent prompts, design workflows. After that, usage feels like any normal tool — but the tool you’re using is one you (partially) built.
Connect your own models
StellarBase is model-agnostic. You can:
- Use the open-source models on StellarCloud (cheapest, EU-hosted)
- Use commercial LLMs via StellarGate (GPT-4, Claude, Gemini — with PII anonymized before they see it)
- Plug in your own models — HuggingFace endpoints, REST APIs, gRPC, Python scripts
- Mix all three — different agents can use different models
Your organization’s custom ML — fraud detector, artifact classifier, failure predictor — becomes a callable tool that any agent or workflow can invoke.
Everything cited, nothing invented
The platform is built so that every output can be traced to an input. Agents cite the specific passage they’re quoting. Workflows produce structured outputs (CSV, JSON, database rows) with source metadata attached to every row. This matters for regulated work — audit trails, compliance reports, court submissions — and it matters for trust in general.
Deploy where the data lives
You’re not forced into anyone’s cloud. Managed EU cloud is the fastest path. On-premise deployment inside your own Kubernetes works identically. For classified or regulated workloads, air-gapped deployments are first-class — the same platform runs with zero internet egress.
Glossary
| Term | Meaning |
|---|---|
| Base | An isolated container — its own data, its own agents, its own permissions. Think of it like a workspace in Notion or Slack. A ministry might have one per department; a law firm one per matter. |
| Connector | A configured link to an external source (Drive folder, SharePoint library, Postgres DB, custom API). |
| DSM | Dynamic Semantic Module — the proprietary engine that builds the knowledge graph linking documents, people, entities, concepts. |
| Agent | A configured AI specialist — system prompt + tool allowlist + knowledge scope + output schema. |
| Tool | Anything an agent can call — built-in (search, OCR), external (your API), or workflow-composed. |
| Workflow | A graph of nodes producing deterministic outputs. Can be triggered on schedule, on event, or on demand. |
| Citation | A back-reference from an AI output to its source — document + page + passage span. |
| StellarGate | The privacy-proxy product. Anonymizes prompts before they reach external LLMs, de-anonymizes the response. |
| StellarCloud | The inference-API product. EU-hosted open-source models available over HTTP. |
Next steps
- Quick Start — connect a source, spin up an agent, run a query
- Platform Overview — deeper tour of the knowledge base + agents + workflows
- Deployment — cloud, self-hosted, air-gapped, hybrid