Skip to main content
Your AI agent’s reply quality is bounded by what it knows. Knowledge is the section of the product that owns what the agent says: where raw content enters the system, how it’s consolidated into a single source of truth, how conflicts and gaps are surfaced for review, and how it’s scoped per agent so different agents have different knowledge bases.
Read this page first if you’re setting up an agent for the first time, then drill into each subsection.

​
The model

Fini’s knowledge model is built around one principle: Articles is the only source of truth. Everything else becomes candidate knowledge through the review / publish workflow that feeds it. Three things are doing the work here:
1

Origins are equal

Sources, Magic Articles, past conversations, Inbox feedback, and background AI signals are all ways new knowledge enters Fini. None of them is the source of truth on its own. A Zendesk article in your Sources is a candidate for knowledge, not the canonical answer.
2

Review Queue is the consolidation layer

Most new knowledge lands here first. Review’s job isn’t only “approve / reject”, it’s to surface gaps (questions customers asked that no article answers) and conflicts (two inputs that disagree) before anything goes live. Human-created drafts can also publish directly into the Published tab; background AI drafts always go to Review.
3

Articles is the knowledge graph

Approved and published knowledge lives in Articles, organized in folders, scoped per agent. This is what the agent retrieves from at runtime. If it’s not in Articles, the agent doesn’t treat it as authoritative.

​
Why Sources can’t be the source of truth

A help center is an artifact written for humans browsing a website. It has marketing language, outdated pages, contradictions across articles, and gaps where the answer is “ask support.” Pointing an agent directly at a help center inherits all of that.
Articles solves this by being the curated, deduplicated, conflict-resolved version of everything your team knows. Sources contribute to it. They don’t replace it.

​
The four pages

Sources

Connect links, files, Google Drive, Notion, Confluence, and Zendesk help centers as raw inputs.

Magic Articles

Paste raw content (transcripts, docs, notes) and generate articles from it — either straight to live or routed through Review.

Review Queue

The consolidation layer. Surfaces gaps and conflicts from every origin before anything reaches Articles.

Articles

The knowledge graph itself. Hosts every approved article, organized in folders, with agent-specific assignment and attribute filters.

​
Agent-specific scoping

Articles is where agent-specific scoping happens. The top-right selector lets you toggle which folders and articles each agent can retrieve from. The same underlying knowledge graph can power very different agents:

Support agent

Account · Billing · Product folders

Sales agent

Pricing · Comparisons · Integrations folders

German-speaking agent

Only German-language folders
If two agents need different knowledge bases, scope at the Articles layer. Don’t duplicate Sources or split workspaces.

​
First-day setup

If you’re starting from zero:
1

Ingest your help center under Sources

Use Sitemap mode if available — it’s faster, more accurate, and gives you the publisher’s canonical URL list.
2

Run bulk Magic Articles generation

Generate over those Sources to draft a structured knowledge base in one pass.
3

Open Review Queue

Read the drafts, the flagged duplicates, the conflict surfaces. Edit, approve, reject.
4

Open Articles

Organize approved articles into folders that map to your customer journey. Use the top-right selector to assign folders to agents.
5

Test in Inbox or Playground

Confirm the agent retrieves the right articles before exposing it to real customers.
Each step has its own page; cross-link freely while you build.