Knowledge Base
The knowledge base is where your content lives — the documentation, policies, product guides, FAQs, and internal runbooks that your AI bots search when answering customer questions and that your agents browse when they need a quick reference. This guide covers creating and organizing collections, uploading content, scoping collections to specific bots, maintaining content quality over time, and using knowledge base feedback signals to improve bot performance.
Overview
The RapiDesq knowledge base serves three audiences:
- AI bots searching for relevant content to ground their answers. Every bot response cites the article it drew from, giving customers verifiable sources and giving your team full audit.
- Agents looking up information mid-conversation. The same content the bot uses is browsable from the agent workspace so human responses stay consistent with bot responses.
- Customers, when you choose to expose the knowledge base as a public help center (optional and configurable per collection).
Content is organized into collections — named groupings of articles that you can scope bots to. A billing bot attached only to the billing collection searches only billing content; a general support bot might be attached to multiple collections for broader coverage. This scoping is what lets you run focused, high-quality bots rather than one mega-bot with a giant undifferentiated content blob.
Collections
A collection is a named group of related articles. To create one, navigate to Admin > AI > Knowledge Base > Collections and click Create Collection.
| Field | Description |
|---|---|
| Name | An internal name for the collection (e.g., "Billing & Payments", "Product Tutorials", "Security & Compliance"). Shown in admin views and in bot configuration. |
| Description | A short description of what's in the collection. Useful for admins choosing which collections to scope a bot to. |
| Visibility | Whether the collection is private (used only by AI bots and agents) or published as part of a customer-facing help center. You can change this later without losing content. |
| Locale | The primary language of the content. Used for search tuning and for matching bots serving specific languages. A collection can have content in multiple languages, but a primary locale helps the system route queries appropriately. |
Common collection patterns
- By topic — "Billing", "Product", "Security", "Getting Started". This is the most common pattern and works well when different bots handle different topics.
- By audience — "Customer-facing", "Partner-facing", "Internal-only". Useful when the same topic needs to be documented differently for different audiences.
- By product or service line — "Platform A", "Platform B", "Professional Services". Useful for multi-product companies where content doesn't cross between products.
- By language — "English", "Spanish", "French". When you have significant multilingual content, separating by language simplifies search quality and bot configuration.
Too many tiny collections become a management burden and make it hard to scope bots cleanly. A good rule: if a bot would need access to more than 3–4 collections, you probably have too many. Aim for 5–15 collections total across the whole tenant, each covering a meaningful topic area.
Adding Content
Content enters the knowledge base in several ways, each suited to different sources.
Direct authoring
The built-in editor lets you write articles directly in RapiDesq. Each article has a title, body (rich text with headings, lists, tables, images, and links), metadata (tags, category, last-updated date), and an optional "verified by" field showing who last reviewed the content for accuracy.
Use direct authoring when content lives natively in RapiDesq and doesn't exist elsewhere — typical for customer-facing help articles you create specifically for support.
File upload
Upload existing documentation in common formats. Supported at launch:
- Markdown (.md) — including GitHub-flavored Markdown with tables and code blocks
- PDF (.pdf) — text-based PDFs are fully supported; scanned/image-only PDFs require OCR preprocessing
- Plain text (.txt)
- HTML (.html) — including documentation exported from other help desk platforms
- Word documents (.docx)
Uploaded files are automatically parsed into articles. Headings become article boundaries when the document contains multiple topics; a single-topic document becomes one article. You can review and split or merge the resulting articles before they go live.
Bulk import
For large-scale migrations from existing help centers or documentation platforms, bulk import accepts a structured archive (typically a zip of Markdown files with front-matter for metadata, or an export from a supported platform). Bulk import preserves structure, metadata, and internal cross-links; broken external links are flagged for review.
API ingestion
For content that's authored and maintained outside RapiDesq (for example, in a Git-backed documentation system, a Notion workspace, or a CMS), the knowledge base API lets you keep RapiDesq's copy in sync. A scheduled job or a webhook from the source system pushes updates; RapiDesq replaces the corresponding article content while preserving its identity, ratings, and attached bot associations.
This is the right pattern when:
- Documentation is source-of-truth somewhere else and you don't want to duplicate editing
- Engineering, product, or marketing teams own content upstream and you want to inherit their changes automatically
- Versioning and history need to live in your existing system (Git, Confluence, Notion) rather than RapiDesq
Article Structure
Every article has a small set of fields that affect how bots and agents find and use it.
| Field | Purpose |
|---|---|
| Title | Short, specific description of what the article covers. Good titles are searchable and unambiguous: "How to update your billing email" beats "Billing" every time. |
| Summary | One or two sentences summarizing the article. Surfaces in bot citations and in search results. Worth writing intentionally rather than auto-generating. |
| Body | The full article content. Supports rich text, lists, tables, images, code blocks, and internal links to other articles. |
| Tags | Free-form labels used for filtering and organization. Examples: "billing", "enterprise", "common-issue", "troubleshooting". |
| Last verified | A date and optionally a user, indicating when the content was last reviewed for accuracy. Articles that haven't been verified recently are flagged in quality reviews. |
| Status | Draft, Published, or Archived. Only Published articles are searchable by bots and agents. Archived articles are kept for history but excluded from search. |
Scoping Content to Bots
Every AI bot is attached to one or more collections. The bot can only answer from content in those collections. Scoping is configured in the bot's Knowledge Base section.
Scoping patterns
- Narrow bot, single collection. The bot searches only one collection. Best for highly specialized bots — a billing bot on the "Billing" collection, a security bot on the "Security & Compliance" collection.
- Bot with multiple related collections. The bot searches a handful of related collections. Useful when a topic area naturally spans multiple collections — a technical support bot with access to "Product", "Troubleshooting", and "Common Errors".
- General-purpose bot, broad scope. The bot has access to most or all customer-facing collections. Used as a catch-all fallback, often combined with more specialized bots that get first crack at specific topic areas.
Scope tradeoffs
Narrower scope produces higher quality answers on the topics covered but refuses more out-of-topic questions. Broader scope answers more questions but with less precision. The sweet spot is usually a small number of specialized bots (narrow scope) combined with one general-purpose bot (broad scope) as a fallback.
When customers report that a bot is giving bad answers, don't just blame the bot or the content — look at the scope. A bot scoped too broadly gets confused by irrelevant content. A bot scoped too narrowly refuses legitimate questions. Adjusting scope is often the fastest quality fix.
Maintaining the Knowledge Base
A knowledge base without maintenance decays. Products change, policies evolve, new issues emerge, and stale content actively hurts bot quality. Plan for ongoing maintenance from the start.
Regular reviews
Set a cadence for reviewing the knowledge base — typically monthly for high-traffic content, quarterly for everything else. The admin dashboard surfaces:
- Articles that haven't been verified in 90+ days — candidates for fresh review
- Articles with low confidence bot citations — the bot uses them but customers rate the resulting responses poorly, suggesting the content may be outdated or unclear
- Articles with no recent bot or agent use — possibly candidates for archival if they're no longer relevant
- Broken internal links — references to articles that have been archived or renamed
Feedback signals
Every bot response includes thumbs-up and thumbs-down feedback from customers, plus a mechanism for customers to say what was wrong. These signals aggregate per article:
- An article that consistently gets thumbs-down when cited probably needs updating
- An article that gets "unclear" feedback needs rewriting for clarity
- A topic area where customers repeatedly indicate "not what I was asking about" may need new articles to fill gaps
Identifying content gaps
The knowledge base dashboard tracks questions the bot escalated because it couldn't find relevant content. These are your content gaps. Patterns of questions without matching articles tell you exactly what to write next. "We keep getting questions about import failures but have no article about them" is a concrete content project with known demand.
Updating articles
When you update an article, RapiDesq preserves its identity (the URL, the bot citations, the history) while updating the content. Pending conversations don't change — if a bot cited the old content five minutes ago, that citation still points to what the bot actually used. New conversations after the update use the new content.
Optional: Public Help Center
Collections with their visibility set to public can be published as a customer-facing help center at a configurable URL (typically help.yourcompany.com or a subdirectory of your main site). The help center shows articles organized by collection, with search, and styled to match your brand. Customers can browse or search without authenticating.
A public help center complements the AI bot — it gives customers a self-service option before they start a conversation, and it gives the bot content a public URL that can be shared in replies and cited in conversations. Exposure is optional; many RapiDesq customers keep their knowledge base private and use it only for the AI bots and agents.
Best Practices
Write for answering questions, not for presenting
A knowledge base article that the AI bot uses well looks different from a marketing landing page. Short, direct, one topic per article. Use concrete headings that match the questions customers actually ask ("How do I reset my password?" beats "Password Management"). Put the answer in the first paragraph and add detail below for readers who want more.
One topic per article
An article that covers three loosely-related topics will be cited for all three but answer none of them cleanly. Split multi-topic articles into focused single-topic articles. Link between them when they're related.
Keep articles current or archive them
An outdated article with the wrong information is worse than no article at all — it gives the bot misleading content to ground answers on. If you can't keep an article current, archive it. The bot won't cite archived content.
Match collection structure to how you want to scope bots
Collections are the unit of bot scoping. Organize content into collections that reflect how you want bots to access it. If you want a billing-only bot, make sure billing content is a single well-defined collection, not scattered across a dozen.
Review content gaps weekly
The list of questions the bot couldn't find answers for is the most valuable input you'll have on what to write next. Review it weekly, even for 15 minutes, and file a ticket or task for the top 1–3 gaps. Over a few months this compounds into dramatically better bot coverage.
Lean on citations, don't hide them
The bot cites its sources by default. Don't configure it to hide that. Customers trust answers more when they can click through to the source, and your team can audit every answer against its source. Hiding citations would be a short-term perception gain and a long-term trust cost.
Related Topics
- AI Bot Configuration — how to configure the bot that searches the knowledge base, including instruction writing, guardrails, and escalation rules.
- AI Credits & Billing — how bot usage is metered and how to monitor credit consumption.
- Conversation Flows — using bots in multi-step flows, including context passing and escalation routing.