The AI knowledge base that answers, not just stores
Most knowledge bases are shelves: articles go in, and finding the right one is the customer's problem. HelpShelf turns the same content into an answering system — hybrid search finds it, trust tiers rank it, and every confident answer arrives with citations.
What is an AI knowledge base?
An AI knowledge base stores your documentation and answers questions from it directly. Instead of returning a list of possibly relevant articles, it retrieves the best passages using combined keyword and semantic search, generates a direct answer, and cites its sources — so users get the answer and the receipt in one step.
The knowledge base nobody reads
You wrote the docs. The answer is in there. And the support requests keep coming anyway — because a traditional knowledge base makes the customer do the retrieval. They have to guess your vocabulary, pick the right article from a results list, and read until they find the paragraph that applies to them. Most will not. They ask a human instead, or worse, they leave.
Keyword search makes it worse: a customer types "my card got declined" and the article is titled "Resolving payment failures." Zero overlap, zero results, one avoidable ticket. The content was never the problem. The interface was.
Retrieval that finds it — however it's phrased
HelpShelf runs hybrid search on every question: BM25 full-text search catches exact terms — error codes, feature names, SKUs — while vector embeddings catch meaning, so "card got declined" still lands on "payment failures." The two rankings are fused, because each one alone fails in ways the other covers.
Keyword + semantic, fused
BM25 and vector similarity run together and merge into one ranking, so exact matches and paraphrases both surface.
Trust-tiered ranking
Curated (publisher-authored or imported) content is boosted above standard (crawled original) pages; AI drafts are deprioritized — and excluded from sales answers entirely.
Citations on every confident answer
Confident answers carry numbered citations back to source articles. If the sources are too weak to cite, the agent says so; with tested Telegram reachability, it can offer a visitor-confirmed handoff.
The trust tiers matter more than they sound. Every knowledge base accumulates three kinds of content — articles a human wrote or approved, pages a crawler pulled in, and drafts an AI generated. Treating them as equals is how bots end up confidently citing their own guesses. HelpShelf ranks curated above standard, penalizes generated content in retrieval, and shows the tier on each citation so you always know what an answer is standing on.
One knowledge base, every surface
Write it once; answer everywhere. The same trust-tiered content powers four surfaces.
The widget on your site
Visitors ask in plain language and get a direct, cited answer — with the source articles one tap away. Support and presale questions alike.
Your email inbox
The same published knowledge powers supported widget answers, hosted help-center content, llms.txt, and four read-only MCP tools.
A public help center
Your curated articles published as a fast, searchable help center — the traditional surface, still there for the people who like to browse.
AI assistants, via llms.txt + MCP
Every site publishes llms.txt and its own MCP server, so ChatGPT, Claude, and Perplexity can query your knowledge base directly and answer about your product accurately.
Built from what you already have
Nobody starts from a blank page. Four ways in, mix and match.
Scan your website
Point HelpShelf at your site and it crawls your pages into searchable, answerable content — classified as standard-tier until a human reviews it.
Build from your codebase
Run npx helpshelf in your repo and the CLI reads your project — README, docs, product surface — and drafts the knowledge base from what the code actually does.
Import from your help desk
One-click imports from Zendesk, Intercom, and Freshdesk bring your existing articles over — marked curated, the highest trust tier, because humans already reviewed them.
Write and curate
A clean editor for the articles that matter most. Anything you write or approve is curated-tier and ranks above everything else in retrieval.
And it tells you what's missing
Every question the agent could not answer well is logged as a knowledge gap. HelpShelf clusters near-duplicate questions, surfaces the gaps on your dashboard, and can draft an article to fill each one — clearly marked as an AI draft and held unpublished until you review it. The knowledge base gets better exactly where customers proved it was weak.
Keep exploring
AI customer support agent
The full guide: how agents work, what to look for, and how to evaluate them.
AI help desk
What happens after an unsupported widget answer — a visitor-controlled handoff through tested Telegram.
Answers that sell
The same knowledge base answering presale questions and capturing leads.
Buying guide
Categories, evaluation checklist, and pricing models across the market.
AEO & GEO guide
How to get your content cited by ChatGPT, Claude, and Perplexity.
For AI agents
The agent-readable surface: llms.txt, context bundles, and the MCP server.
Frequently asked questions
An AI knowledge base stores your documentation and answers questions from it directly. Instead of returning a list of maybe-relevant articles, it retrieves the best passages using keyword and semantic search combined, generates a direct answer, and cites the source articles — so users get the answer and the receipt in one step.
A normal knowledge base is a filing cabinet: it stores articles and hopes users find and read the right one. An AI knowledge base is an answering system: it understands the question, pulls the relevant passages even when the wording does not match, and responds directly with citations. The articles still exist — they become the evidence rather than the interface.
Three mechanisms. Hybrid retrieval (BM25 keyword search plus vector similarity, fused) finds the genuinely relevant passages. Trust tiers rank publisher-authored and imported content above crawled original pages and deprioritize AI drafts — which are excluded from sales answers entirely. And citations on every confident answer make each claim checkable; when confidence is low, the agent says so and can offer a visitor-confirmed handoff only when tested Telegram is reachable.
Yes. HelpShelf imports articles from Zendesk, Intercom, and Freshdesk with their structure intact, and can crawl any public help center or website. Imported articles are marked curated — the top trust tier — since a human already reviewed them, and re-imports update existing articles rather than duplicating them.
Yes — that is a core design goal. Every HelpShelf site publishes llms.txt (a map AI tools read), a full context bundle, and its own MCP server with search and ask tools. When a prospect asks ChatGPT or Claude about your product, those assistants can query your real documentation instead of improvising from stale training data.
HelpShelf has a free plan with no credit card required — enough to build the knowledge base, publish a help center, and put the widget live. Free includes 500 AI searches and 100 AI chats each month. Pro is $49/month for 10,000 searches and 2,000 chats; Scale is $149/month for 100,000 searches and 20,000 chats.