CRMown Docs
Docs › AI Features

Knowledge Base

Org-shared answers the AI consults BEFORE making a model call. Curated by your team — promoted from AI Bar History or written manually. The longer you use the platform, the smarter it gets, and the more your AI calls hit the KB cache instead of paying for new tokens (regardless of which AI provider your org runs on).

Open in CRMown →

Steps

  1. Sidebar > Tools > AI Intelligence > Knowledge Base. Browse, search, or filter articles by category + tag.
  2. "+ New Article" opens a tabbed dialog with three add-modes:
  3. • Manual — title, question, answer, category, tags. Same as before; required for entries that don't come from another source.
  4. • From URL — paste a public URL (HTML page or PDF). The backend fetches it, strips boilerplate, and the CRMown AI chunks the content into up to 12 atomic Q/A articles. Each article is saved with sourceType="imported" and sourceId pointing at the URL.
  5. • From File — drag/drop or pick a PDF, DOCX, MD, or TXT (5MB max). Same chunk-with-AI pipeline. Articles are saved with the original filename as sourceId.
  6. Promote any AI Bar answer via the "↑ Save to Knowledge Base" pill on the response bubble itself, or via the "↑KB" button on the AI History row for older entries. Both paths create a KbArticle with sourceType="ai_bar" and sourceId pointing back to the AiBarHistory row, then back-link the history row (isPromoted=true, kbArticleId set). Re-promoting is idempotent — returns the existing article instead of duplicating.
  7. Edit / delete buttons are gated to creators + admins. Anyone in the org can read.
  8. Behind the scenes: every AI Bar query first calls findRelevantArticles(orgId, question) — if a high-similarity match exists, the answer is returned from KB without a model call (saves tokens on any provider, faster, deterministic).
  9. Feedback buttons (👍 / 👎) on each article train the relevance ranker over time.

Good to know

  • KB hit rate compounds — orgs with 50+ articles see 30-40% of AI Bar queries answered from KB (zero AI cost on any provider). Worth investing the curation time early.
  • Best candidates for KB: recurring questions ("how do I export contacts?", "what's our refund policy?"), domain knowledge ("our pricing tiers"), troubleshooting ("when X happens, do Y").
  • KB articles are org-scoped — competitors trialing the platform can't see your knowledge base, and your KB doesn't leak into other orgs'.
  • Search uses both semantic similarity (for "what does our docs say about pricing?") and exact tag/category match (for "show me all billing articles").
  • Ingest cost: each URL/File ingest is one standard-tier AI call regardless of how many Q/A pairs come out. A typical 10-page docs page or PDF lands at well under 5K input tokens. The 200KB extracted-text cap (≈ 30-40 pages) keeps any single ingest from running away with tokens.
  • Best ingest sources: existing FAQ pages, product documentation, internal wikis exported to PDF/DOCX, customer-facing help center articles. The chunker skips boilerplate (footers, cookie banners, copyright) automatically.
  • If an ingest finishes with 0 articles created, the source had no usable knowledge content — the modal shows that as the run's error. Common cause: pasting a marketing landing page that's mostly hero images and CTAs.
Example: A 40-person agency promotes 80 questions to KB over their first month. By month 3, their AI Bar token usage drops 35% even though their team's AI usage doubled — KB is absorbing the FAQ traffic and only novel questions hit the AI provider. Their bill stays flat as their team grows.

Troubleshooting

AI Bar still uses tokens for questions I know are in the KB
Either the question phrasing didn't match closely enough to trigger the KB cache, or the KB article was deleted / unpublished.
Fix: Check Knowledge Base for the matching article. Rewording the question or marking the article with more relevant tags / synonyms improves hit rate.
Can't edit an article
You're not the creator and you're not an admin. Edit access is gated.
Fix: Ask an admin to edit, or to grant you the article's edit permission.
KB feedback buttons don't respond
AI quota is exhausted (the feedback ranker uses the light tier) or the article doesn't have a feedback row yet.
Fix: Check Settings > AI Usage. Click the article first to load it, then submit feedback.
URL ingest fails with "fetch failed"
The page requires authentication, blocks bots, or returned a non-2xx status.
Fix: Use a public URL, or download the page and use the From File tab instead.
File ingest rejects a PDF that opens fine in a viewer
PDF is image-only (scanned) — text extraction returns nothing. Or the file exceeds the 5MB cap.
Fix: Run the PDF through OCR first (e.g. Adobe Acrobat's "recognize text"), then re-upload. For oversize files, split into smaller PDFs.

Plans

Included in every plan. KB itself is free; the AI Bar pre-check uses a tiny similarity lookup that doesn't make any model call. Toggleable per-org under Settings > AI Usage > AI Features (key: knowledge_base) — disabling it stops the pre-check (AI Bar still works without it, just calls the model every time).

Related

  • AI Bar History (the source of most KB articles via "Promote to KB")
  • AI Command Bar (Ctrl+K) (the consumer of KB articles — KB cache fires before any model call)
  • Training Articles (this corpus, which the SOS helper uses — distinct from org-curated KB)

View as Markdown