HERC: your AI assistant and home page¶

HERC is the home page of Datahub and a conversational assistant that answers your questions, summarises your work, and — on request — acts across the platform for you, always inside your own roles.

Describe what you want and HERC takes you there — so you never have to learn the platform's information architecture before getting value from it.

SurfacesHome page · chat on every page

RoutingFastRouter · EN · NL

ScopeYour JWT · your roles

Accessherc.read — every user

When to choose this¶

Find something fast

"Show me the Customer Lifetime Value metric." → HERC links you straight to it.
Ask a business question

"What was revenue last month?" → HERC routes through the metric agent and answers with the number.
Summarise a body of work

"What changed in the Finance domain this week?" → HERC reads activity logs and gives you a digest.
Get suggestions

"Suggest five metrics for the new Subscriptions contract." → HERC analyses the contract's tables and proposes them.
Handle a vague request

"Where do I configure provider keys?" → HERC navigates you there.
Hand off to a specialist

A complex catalog or glossary question is routed to the right module agent automatically.

When not to reach for HERC

Bulk operations → use the module's own import / export.
Privileged admin actions HERC doesn't have access to → HERC respects your roles and will tell you when a role is missing.
Free-form analysis on raw warehouse tables → HERC works through governed metrics; for ad-hoc work use a notebook or Insights.

How HERC works¶

HERC is a routed agent system. When you send a message it decides — in milliseconds — whether a specialist can answer directly or whether it should reason about your request first, then it works through governed tools and streams the answer back.

flowchart LR
    Q["Your question<br/>English or Dutch"] --> R{"FastRouter<br/>keyword match"}
    R -->|"strong match"| A["Specialist agent<br/>metric · glossary · catalog · alerts"]
    R -->|"ambiguous"| H(["HERC triage agent<br/>reads your role · page · activity"])
    H --> A
    A <--> G[("Governed sources<br/>Organisation DNA · catalog · metrics")]
    A --> ANS["Grounded answer<br/>links · navigation · follow-ups"]
    ANS -.streamed.-> Q
    classDef accent fill:#ff8400,stroke:#e67700,color:#ffffff,font-weight:600;
    class H accent

Figure 1 — How HERC answers a question. A confident keyword match (English or Dutch) routes straight to a specialist; anything ambiguous goes to the HERC triage agent, which reasons over your context and delegates. The agent reads only governed sources your roles allow, then streams a grounded answer back with links you can click.

The same journey, step by step:

Fast routing. A keyword router inspects your message in English and Dutch. If it strongly matches a module's keywords (metric, kpi, dashboard, glossary term, …) it routes directly to that module's agent.
Reasoning fallback. Otherwise the message goes to the HERC triage agent (the orchestrator), which reads the platform's context — your role, the page you're on, recent activity — and decides which agent or tool to call.
Specialist agent. A module agent (the metric, business glossary, or catalog agent, and others) executes, using that module's governed tools — search assets, get a metric, draft a definition, and so on.
Streamed reply. Tokens stream into the chat panel. Tool calls surface as inline status messages ("Searching the glossary…").
Follow-ups. After a reply, HERC suggests two or three follow-up questions you can click to continue.

Why teams reach for this

You describe intent in plain language — English or Dutch — and HERC finds the metric, term, table, or page for you, or hands off to the specialist that can. No menu-spelunking, no query language, and nothing to memorise about where things live.

How HERC stays in your lane

Every call HERC makes is scoped to your JWT, and the same role checks that gate the UI gate the agent. HERC cannot see, modify, or surface anything you don't already have a role to see — so it is safe to let it search on your behalf.

Agentic loop — plan, act, confirm¶

For multi-step tasks (for example, "create three metrics for the Subscriptions contract") HERC shows its work before and during execution.

flowchart LR
    START["Multi-step request"] --> PLAN["Create a plan<br/>numbered checklist"]
    PLAN --> STEP["Run next step<br/>pending → in progress → done"]
    STEP --> CLEAR{"Clear how<br/>to proceed?"}
    CLEAR -->|"yes"| MORE{"More steps?"}
    CLEAR -->|"genuine fork"| ASK(["Pause and ask you<br/>quick-reply chips"])
    ASK -->|"your answer"| STEP
    STEP -.->|"revise plan · max 2 per turn"| PLAN
    MORE -->|"yes"| STEP
    MORE -->|"no"| DONE["Collapse to summary<br/>Thought for 6 s · 4 steps"]
    classDef accent fill:#ff8400,stroke:#e67700,color:#ffffff,font-weight:600;
    class ASK accent

Figure 2 — The agentic loop for multi-step work. HERC plans, runs each step while its live line shows what it is doing, and revises mid-turn (at most twice) when it learns something new. It only interrupts you at a genuine fork — otherwise it runs straight to a one-line summary.

What you see¶

While HERC works, it shows what it is doing above its reply:

One live line — the step HERC is on right now ("Running Spark SQL query on Databricks…"), with the time so far. It changes in place as HERC moves on.
The steps it already took — each with what came back, for example "3 rows · year, total_weight" under the query that returned them, or the error HERC then corrected. Older steps fold into "2 earlier steps".
Plan progress — when HERC made a plan for the work, its progress ("2/3 steps") sits beside the live line. If HERC discovers something unexpected (a table name has changed), it revises the plan; revisions are capped at two per turn.
Completion summary — once the turn completes, this collapses to a one-liner such as "Thought for 6 s · 4 steps". Open it to see every step, what each returned, and the plan. When you reopen a conversation later, it shows the tools HERC ran and what each returned.

Each turn names its author at the top: HERC in its accent colour, and You over your own messages.

Asking for your input¶

If HERC reaches a genuine fork — for example it found two tables that could both be the right source and neither is obviously correct — it pauses and shows a clarifying question with optional quick-reply chips:

"Which table should I use as the base for the Active Customers metric — customers_raw or customers_deduplicated?"
[customers_raw] [customers_deduplicated] [Show me both]

Tap a chip or type your own answer. HERC resumes from exactly where it paused — no need to repeat yourself.

Scroll-back¶

Plans persist with the assistant message. If you scroll back through a conversation you can see what HERC planned and which steps it took, in order.

When HERC pauses vs presses on¶

Situation	What HERC does
Discovers a minor schema difference	Updates the plan silently, continues
Encounters a clear ambiguity tools cannot resolve	Pauses and asks you
Hits a tool error it can recover from	Marks the step failed, revises, retries
Hits a permissions error	Stops and explains what role is needed

What HERC looks like¶

Surface	Where	What you see
Home page	`/`	Greeting, trending entities (top assets / metrics / terms by recent activity), recent activity feed, pending tasks count, the chat input.
Chat panel	Bottom-right of every page	Click the orange spark icon → side panel opens with conversational chat. Persistent across pages.
Conversation history	Chat panel → history icon	Past conversations, restorable.
Suggestions row	Below the chat input	One-click prompts contextual to where you are (e.g. on a metric detail: Explain this metric, Show similar metrics).

What's on the home page¶

The home page is HERC's "what's happening" dashboard:

Trending entities. Top entities by activity count over a configurable window (last 7 / 30 days). Calculated by aggregating audit-log events across catalog assets, glossary terms, metrics, contracts, products, processes, alert rules, and flows. Click any row to jump to the entity.
Recent activity. Stream of platform events: assets updated, terms approved, contracts published, alert rules triggered. Tells you what your colleagues have been doing.
Pending tasks. Count of open inbox tasks (the actual list lives in Tasks).
Quickstart panel. For new tenants: step-by-step on-ramps to each module ("Connect your warehouse", "Define your first metric", "Build a process").

Setup — what an admin needs to do once¶

Prereq	Where	Why
AI provider key	`/admin/integrations` → AI Provider Keys (or `/ai/provider-keys`)	Required for all conversational AI. Without it HERC can answer simple navigation questions but can't reason.
Roles	`/rolegroups`	`herc.read` is included in every authenticated role group by default — every user sees HERC.
Brain index built	`/dna`	The first time a tenant logs in, HERC's answers improve dramatically once the Organisation DNA graph has been built.
(Optional) Private AI	License + `/admin/integrations`	Tenants on the Private AI license can run HERC on the platform's hosted model — see the AI platform page.

Examples worth trying¶

You ask	What happens
"How many active customers do we have?"	Routed to the metric agent → searches metrics → finds Active Customers → executes the metric → returns the number with a link to the metric.
"What's the definition of churn?"	Routed to the glossary agent → returns the Churn term's definition with a link.
"Find me PII tables in the warehouse."	Routed to the catalog agent → semantic search across tagged assets → returns top matches.
"Summarise the changes to the Finance domain this week."	Orchestrator → reads audit logs filtered by domain tag → returns a bulleted digest.
"Set up a new metric for Net Revenue."	Orchestrator → metric agent proposes a draft (table, measure, time grain) → you accept / edit.
"What's broken right now?"	Routed to the alert agent → returns currently-firing alert rules.
"Set a Growth goal for our weakest pillar."	HERC asks for whatever you haven't given (target stage, score, deadline), repeats the goal back, and sets it only once you agree.
"I need access to edit glossary terms."	HERC names the role you're missing → asks for your reason in your own words and confirms the role, reason and duration → files a request that someone else approves. It never grants access itself, and if the role can only be assigned by an administrator it says so instead.

Limitations¶

Limit	Why	Workaround
HERC can't run SQL you write directly.	The platform routes through governed metrics and catalog assets.	Use Insights for ad-hoc questions; HERC will help you build the card.
HERC respects your roles strictly.	A user without `metric.read` cannot ask HERC for a metric value.	Have an admin grant the role.
Some tools are advisory (suggestions), not mutating.	Safety — HERC suggests; you approve.	Accept the suggestion to apply it.
Long conversations may go stale.	Provider context limits.	Start a new conversation when switching topics.
Brand-new tenants get thin answers.	The DNA graph hasn't been built yet.	Build the brain index from `/dna`.
HERC can't see audio or images.	Text-in / text-out.	For meeting notes, use Transcription and ask HERC about the transcript.

Audit & compliance¶

Question a CISO might ask	Where to look
"Who asked HERC what?"	All conversations are stored per-user; admins can audit via `/ai/observability` (Arize Phoenix).
"Did HERC see data this user shouldn't?"	Tool calls are scoped to the user's JWT; the same role checks that gate the UI gate the agent.
"Where do queries physically go?"	Whichever AI provider is configured under `/admin/integrations`. With Private AI, queries stay on the platform.
"Can HERC make irreversible changes?"	Mutating tools require explicit confirmation in the UI. HERC proposes; you click apply.
"Are conversations retained?"	Yes, per-user, with configurable retention. Archive any conversation from the history panel.

Troubleshooting¶

Symptom	Likely cause	Fix
HERC says "I don't have a provider configured"	No AI provider key.	`/admin/integrations` → AI Provider Keys → add one.
Answers feel generic / wrong	Brain index hasn't been built or is stale.	`/dna` → re-index.
HERC can't find a metric you know exists	The metric is archived or you don't have `metric.read`.	Check archive state; verify your roles.
Trending list is empty	Tenant is new — no activity yet.	Use the platform for a few days; trending populates from audit logs.
Chat panel stuck on "Thinking…"	Provider rate-limit or network blip.	Cancel and retry; check `/ai/observability` for the trace.
Follow-up suggestions are missing	The agent that answered doesn't have follow-ups configured for that tool.	Cosmetic; doesn't affect the answer.