Fini (usefini.com) treats your raw documents as candidates, not answers: the agent retrieves only from approved Articles, and Fini’s background AI continuously flags conflicts (two inputs that disagree) and gaps (questions no article answers) in the Review Queue for a human to resolve. Keep generation on Suggest for Review and messy, outdated, or contradictory content is caught at review time, before your agents see it. This page explains which mechanisms catch which problems, how scoping keeps the wrong content away from the wrong agent, and a practical workflow for cleaning up a knowledge base that has grown messy.

​
Why messy docs are a risk

A help center is written for humans browsing a website. It usually has marketing language, outdated pages, contradictions between articles, and gaps where the real answer is “ask support”. Internal docs add drafts, duplicates, and policy that changed but was never cleaned up. Pointing an agent directly at all of that means the agent inherits every contradiction. Fini’s knowledge model is built around one rule: Articles is the only source of truth. Sources, Magic Articles, past conversations, and Inbox feedback all produce candidate knowledge. Most candidates reach Articles through the Review Queue, where conflicts and gaps are resolved first. Drafts generated with Live status, or created while the workspace’s require-review setting is off, skip Review and publish directly, so they get no conflict check.

​
How problems get detected

Different mechanisms catch different kinds of mess. Each one lands its findings in the same place, the Review Queue, or in Analytics when it shows up in live traffic.
Drafts created by Fini’s background AI always land in In Review, never directly in Published, even if your workspace has turned off the require-review setting. Conflict and gap detections are flagged for human judgment, never auto-published.
How a flagged problem moves from detection to a fixed article:

​
How Fini reads an approved article

Fini’s architecture is described as RAGless: the model reads whole approved articles rather than retrieving fragments of documents. That matters for messy content. A condition buried in the fifth paragraph of an article, such as “this applies only to accounts opened before 2024”, stays attached to the answer it qualifies instead of being split off into a separate fragment. See What is RAGless for the architecture. It also means the quality of each article matters more than the quantity. One complete, conflict-free article per topic gives the agent a better answer than five overlapping ones.
To confirm (internal, remove before publish): The product docs say Article Questions are used for “semantic retrieval” and that long uploaded files “are chunked” into documents in Sources, while llms.txt describes RAGless as “no chunking or embeddings”. Confirm the accurate wording: is chunking only a Sources ingestion step, and how are articles matched to a message before being read whole?

​
Which source wins

Precedence between sources is decided by people, at review time:
  • Review decides the canonical answer. When the background AI flags a conflict, it proposes a unified answer. Approve it only if it is correct; otherwise reject it and fix both articles by hand. When two inputs propose competing updates, approve one, reject the other, or merge them in the RESULT column. Don’t approve both.
  • Articles is authoritative. Once approved, the article is what the agent uses. A Source that still says the old thing doesn’t override it, but the background AI keeps flagging the conflict while the source disagrees, so fix or remove the source too.
  • The more specific article usually wins. If two articles both match a message, the more specific one usually wins retrieval. If that isn’t the behavior you want, merge the two through Review.
To confirm (internal, remove before publish): Is there any source-priority or “trusted source” setting (for example, preferring Confluence over the public help center) beyond human resolution in Review? If so, document it here.

​
Scoping keeps the wrong content away

Some apparent conflicts are not conflicts at all: two answers are each correct for a different audience. Scope them instead of merging them.
Don’t duplicate to scope. Copying an article into a second folder so another agent can see it creates exactly the kind of conflict Review exists to catch. Scope at the folder-attachment layer instead.

​
A practical cleanup workflow

Use this when you are onboarding a messy knowledge base, or when Analytics shows Conflicting Knowledge or Missing Knowledge rising.
1

Ingest everything, but promote carefully

Add your help center, internal docs, and files as Sources. Prefer XML Sitemap mode for websites so you get the canonical URL list instead of noise pages. Use Preview content to confirm scanned PDFs actually extracted text.
2

Run a dry run for duplicates

In Magic Articles, run with only Detect Duplicates enabled. No articles are written; Fini only reports which inputs overlap existing articles. This tells you how much consolidation work is ahead.
3

Generate into Review, not Live

Run bulk generation with all three Generation Actions enabled and Suggest for Review selected. Never use Live on content you haven’t vetted.
4

Triage the queue by type

In Review, filter by Type. Clear Duplicate entries first (fastest decisions), then Update entries (read the CURRENT and RESULT diff and use Revert on fields that are wrong), then New entries (confirm the folder). Filter Author to the background AI to work AI-detected conflicts and gaps as their own batch.
5

Fix the source, not just the article

When a conflict came from an outdated source, update or delete that source too. Otherwise the conflict still exists in your Sources, and the background AI will keep re-flagging it.
6

Organize and scope

In Articles, arrange approved articles in folders that match your customer journey, then assign folders to agents and add attribute filters where answers depend on the customer.
7

Verify against real questions

Test in Inbox and with a Test Suite set using the Knowledge consistency judge, which checks whether the agent answered from the correct article.

​
Keep it clean afterwards

  • Work the queue on a schedule. Review has no automatic assignee, so drafts sit until someone opens it. Put a regular owner on it.
  • Scan the Published tab. It is the audit log of everything that reached your agents, including what the background AI proposed and a teammate approved.
  • Refresh sources after copy changes. Sources don’t auto-refresh on a schedule by default. Re-crawl after any substantial help center update.
  • Watch the escalation reasons. In Analytics, a rise in Conflicting Knowledge, Missing Knowledge, or Partially Available is the earliest production signal of knowledge drift. Knowledge performance shows which slices have a low AI Resolve Rate.
  • Turn resolutions into articles. When a human resolves something the agent couldn’t, use Generate Knowledge in Inbox to turn that resolution into an article. Its Status After Generation defaults to Live, so switch it to Suggest for Review if you want a second pair of eyes.

Review Queue

Resolve gaps, conflicts, duplicates, and proposed updates.

Magic Articles

Generate articles with create, update, and duplicate detection.

Sources

Ingest, refresh, and preview raw inputs.

Articles

The source of truth: folders, scoping, and attribute filters.