Documentation
Ultron answers questions from your own documents — and refuses when the evidence isn't there. This guide covers how the governed loop works and how to build on it.
Every answer Ultron produces is grounded in retrieved passages, carries citations, and passes through whatever human review you configure. Nothing is generated from the model's own memory of the world.
Quick start
- Create a knowledge baseUpload the policies, reports and past answers you want Ultron to draw on. Each file is parsed into blocks and embedded so passages can be retrieved individually.
- Pick or build a workflowStart from a template such as the security questionnaire agent, or assemble your own in the visual builder.
- Run itChoose the knowledge base and provide the run's inputs. Each step's inputs and outputs are recorded as it executes.
- Review the draftsApprove, edit or reject each answer. Steps marked as requiring mandatory review will not be considered complete until someone signs off.
- Export and reuseExport the finished packet. Approved answers are retained and become evidence for future runs.
Retrieval quality depends on what is in the knowledge base. If a question has no supporting passage, Ultron refuses rather than guessing — that is the intended behaviour, not a failure.
Core concepts
| Term | What it means |
|---|---|
| Knowledge base | A collection of documents that answers can be grounded in. Also called a vault. |
| Document | An uploaded or imported file, parsed into blocks and embedded for retrieval. |
| Workflow template | A reusable, versioned definition of the steps a run performs. Published versions are immutable. |
| Node | One step on the builder canvas — a source, retrieval, drafting, review or analysis tool. |
| Run | One execution of a workflow against a knowledge base, with recorded inputs and outputs per step. |
| Review item | A drafted answer awaiting a human decision: approve, edit or reject. |
| Citation | A pointer from a claim back to the document, page or approved answer that supports it. |
| Trace event | An inspectable record of a retrieval, generation or verification step. |
Retrieval & grounding
Ultron retrieves with a hybrid of vector similarity and keyword search, fuses the two result sets and reranks them. The combined evidence then passes an answerability gate that classifies the question as answerable, partial or insufficient.
Drafting sees only the retrieved passages — never the whole document set. Afterwards, every sentence is checked back against that evidence. Claims that cannot be supported are not shipped: the answer is either rewritten strictly from evidence or refused.
A confident wrong answer in a compliance reply is far more costly than a gap. Refusals are surfaced as missing information so a person can fill it in deliberately.
Review gates
Review is configured per step and takes one of three values:
- Off — the step completes on its own.
- Optional — a reviewer may weigh in, but the run continues.
- Mandatory — the output is held for explicit sign-off.
Reviewers see the drafted answer, its citations, and any missing information the pipeline flagged, so a decision doesn't require re-reading the source material.
Approved memory
Once a human approves an answer it is stored and becomes retrievable evidence for later runs. The second time a question is asked, the approved answer is available alongside the source documents — which is what makes repeat work progressively faster.
Workflow builder
Workflows are node graphs you edit on a canvas. Drag a node from the palette, connect it to the steps that depend on it, and configure its parameters in the inspector. Published versions are immutable; editing a published workflow creates a draft.
- Sources pull documents in from a connector.
- Retrieval finds supporting evidence for a question.
- Drafting writes answers from that evidence.
- Review creates the items a human signs off on.
- Analysis summarizes and compares documents.
Use Tidy up in the canvas controls to arrange a graph automatically.
Connectors
Connectors import documents from the systems your team already uses into a knowledge base, where the same governed loop applies. Connect a credential once on the Integrations page, then reference it from a source node.
Connector secrets are encrypted at rest. If the encryption key is not configured, Ultron refuses to store a credential rather than saving it in plaintext.
Re-importing the same upstream record adds a new version to the existing document instead of duplicating it, so a source can be refreshed safely.
Examples
Models & providers
The language and embedding models are configurable per organization from the Settings page — no redeploy required. Self-hosted and OpenAI-compatible endpoints are both supported, so evidence never has to leave your network.
Embeddings must be produced by the same provider that embeds queries at retrieval time. After switching, re-index existing knowledge bases so old and new vectors aren't compared.
FAQ
What happens when there is no supporting evidence?
Ultron refuses and reports what was missing. It never fills the gap from the model's general knowledge.
Is any of my document set sent to the model?
Only the top-matching retrieved snippets are included in a prompt — never the whole corpus.
Are agents autonomous?
No. An agent is a fixed, reviewable sequence of known steps. There is no free-form planner deciding what to do next at runtime.
Can I see why an answer was produced?
Yes. Retrieval traces, per-step inputs and outputs, and citation checks are all inspectable for every run.
Still stuck? Get in touch.