Frequently Asked Questions
Everything you need to know about Kelu.
Also useful: all features · alternatives · compare · trust center · free tools.
General
What is Kelu?
Kelu is an AI Knowledge Assistant. It connects your docs, wikis, policies, product content, code repos, community forums, and helpdesk knowledge — then answers questions with citations on every surface: a hosted widget, Slack / Discord / Microsoft Teams bots, a browser extension, an MCP server for AI IDEs, Support Ticket AI inside your helpdesk, and a public API + SDKs.
How is Kelu different from just using ChatGPT?
ChatGPT answers from its training data, which doesn't include your product documentation, internal policies, or private knowledge. Kelu retrieves from your indexed content and cites the exact page each answer came from — so teams and customers can verify the response, and the assistant abstains when it doesn't have grounded evidence rather than making something up.
Do I need machine-learning expertise to use Kelu?
No. You connect a source (a URL, a GitHub repo, a Notion workspace, a helpdesk) and Kelu handles chunking, embedding, indexing, retrieval, reranking, and generation. Nothing to fine-tune, no infrastructure to run.
How does Kelu compare to Kapa / Inkeep / Chatbase / Intercom Fin?
We have detailed side-by-side comparisons at /alternatives (22 competitor pages) and /compare (category-level round-ups). The short version: Kelu ships more surfaces on one plan (widget + Slack + Discord + Teams + MCP + browser extension + Support Ticket AI + form deflector), has a real self-serve free tier, and combines hybrid retrieval (vector + BM25 + reranker) with an enforced abstention gate.
RAG, Citations & Grounding
What is RAG and why does it matter?
RAG (Retrieval-Augmented Generation) retrieves relevant documents first, then generates an answer grounded in those documents. This eliminates hallucinations about your specific product and makes every answer verifiable via citation links. See /glossary/rag for a longer explanation.
How are citations generated?
During retrieval, each source chunk is tracked with its URL and title. The generation prompt requires the model to ground in those sources. Every response includes a structured citations array with the URL, title, and excerpt for every source chunk used — the widget, bots, MCP, and API all return the same citations.
What happens when Kelu can't find a good answer?
When retrieval confidence falls below the configured threshold, the assistant abstains — returning a transparent fallback message instead of hallucinating. This is enforced by an abstention gate, not just prompted. Unanswered questions become part of the gap report and can be turned into Doc Proposals (AI-drafted new pages your team reviews).
How does hybrid search work?
Kelu combines semantic vector search (pgvector cosine) with full-text BM25 ranking. Results are merged with reciprocal rank fusion. Before search, an LLM rewrites the query to improve recall. After retrieval, a cross-encoder reranker sorts the top candidates. See /features/hybrid-search for the full pipeline.
How do you measure answer quality?
LLM-as-judge evals score every answer for groundedness and citation quality; scores trend per knowledge base per week. Combined with user thumbs-up / thumbs-down and the gap report, you see quality as a moving number, not a demo. See /features/evals-quality.
Surfaces & Integrations
Where can Kelu answer questions?
Hosted web widget, Slack / Discord / Microsoft Teams bots (all native), a browser extension (Chrome / Edge / Opera MV3 side panel), an MCP server per knowledge base for Cursor / Claude Desktop / ChatGPT with MCP, a standalone form deflector for support-form ticket deflection, Support Ticket AI inside every major helpdesk, and a public REST + WebSocket API with SDKs for web / React / React Native / Node.
Can Kelu index GitHub repositories?
Yes. Connect a GitHub repository and specify which paths to index (README.md, docs/, ADRs, changelogs). A push webhook triggers incremental re-index on every commit to the configured branch. GitHub issues, pull requests and discussions are separate source types you can index alongside the repository.
Which helpdesks does Support Ticket AI integrate with?
Zendesk, Freshdesk, Front, Intercom, Jira Service Management, Linear, and Salesforce Service Cloud. Each integration surfaces AI copilot drafts inside the helpdesk UI, and can auto-resolve tickets above a per-queue confidence threshold. See /features/support-ticket-ai.
How does ticket deflection work?
Two paths. Before submit: the /features/support-form-agent form deflector answers questions in-line as the customer types into a support form — confidence-scored, and only deflects when it has grounded evidence. In the helpdesk: Support Ticket AI reads incoming tickets and either drafts a reply for the agent or (above your confidence threshold) auto-responds. Simulations let you calibrate the threshold on historical tickets before enabling anything customer-facing.
Can I test Support Ticket AI before enabling it in production?
Yes — that's what Simulations do. Point at your last N days of resolved tickets, the assistant answers each as if it were live, and you review the results in the dashboard with human sign-off per answer. Per-topic accuracy scores and confidence distributions become the artifact you take to a launch decision. See /features/simulations.
What is the MCP server and how do I use it?
Every knowledge base is reachable over Model Context Protocol (MCP) at POST /mcp. Authenticate with that knowledge base's X-Client-Key header, which is what scopes the server to it. Cursor, Claude Desktop, ChatGPT (with MCP), Windsurf, and any MCP-compatible agent can list and invoke search_docs, ask_question, and list_sources as structured tools. See /features/mcp-server.
What's the browser extension?
A Chrome / Edge / Opera MV3 side-panel extension that opens your knowledge base next to whatever tab you're on. Ask questions about the page, look up an API without leaving your editor, or pull an internal policy answer without breaking flow. Signs in with your Kelu account and respects the same permissions as the dashboard. See /features/browser-extension.
How do the Slack / Discord / Teams bots work?
Installed per workspace from the dashboard. Slack: /askai slash command + @mention + forum auto-reply. Discord: mentions + slash commands + forum auto-reply. Teams: mention-based Q&A. All three carry citations, thumbs feedback (up/down), and identical answer quality — same retrieval pipeline as the widget and API.
What are Collections and per-collection permissions?
Collections (source groups) let you scope retrieval by product line, version, or region. Client keys can enforce allowed groups — narrowing what an integration can retrieve, never widening it. A partner-facing key that only sees that partner's product docs is one call to configure.
Does Kelu support versioned documentation?
Yes. Tag sources with a version (v1, v2, latest) and issue client keys scoped to that version. Users always get answers from the docs version that matches their product version.
Doc Proposals
What are Doc Proposals?
Gap analytics logs every unanswered question and ranks the missing knowledge by volume. Doc Proposals drafts the missing page in your tone — your team accepts or dismisses each draft. AI proposes, humans publish. See /features/doc-proposals.
Security, Privacy & Data
Is my content used to train AI models?
No. Indexed content, embeddings, and conversations stay isolated to your workspace and are used only to answer your own questions. Nothing is shared across customers or used to train any model — ours or a third party's. This is contractual on paid plans and technical on every plan.
How is PII handled?
PII (emails, phone numbers, credit-card numbers, common ID formats, personal names) is masked before every LLM call — chat, reranking, evals, deflection, agents — and before storage. The database only holds the masked form. See /features/pii-and-zero-retention.
What is zero-retention mode?
A per-workspace toggle. Questions and answers stream to the caller in real time and are then dropped — no conversation row is created, no message is stored. Retrieval still runs against your indexed content and the answer is still cited. Only the transcript is not persisted.
Can I use my own OpenAI / Anthropic / Gemini key (BYOK)?
Yes on Enterprise. BYOK routes LLM traffic through your provider account with your key, under your existing data-processing agreement. You keep the model choice and the invoice; Kelu is the platform layer. See /features/byok.
Is Kelu SOC 2 certified?
SOC 2 Type II is in progress (Vanta-managed; audit scheduled). HIPAA BAA, GDPR + SCCs, and regional data residency (US / EU) are available on Enterprise. See /trust for the current control posture.
Where is my data stored?
US by default. EU data residency is available on Enterprise. Data is encrypted at rest (AES-256) and in transit (TLS 1.3).
Do you support SSO?
Yes on Enterprise: OIDC and SAML 2.0. Works with Okta, Entra (Azure AD), Google Workspace, OneLogin, and any SAML-compliant IdP.
Is the widget protected against abuse?
Yes. The hosted widget at widget.kelu.dev is protected by reCAPTCHA to prevent automated bots from consuming your question quota. Runs invisibly; no user friction. Can be disabled for authenticated / internal surfaces.
How are client keys secured?
Public client keys (kl_pk_...) are safe to embed in client-side code — they're scoped to a single knowledge base and only allow Chat and Search calls. Backend integrations use a client ID + secret pair (kl_ci_... / kl_cs_...) which must not be exposed client-side. Keys are revocable instantly from the dashboard.
Pricing & Limits
Is there a free plan?
Yes. The Free plan includes 1 knowledge base, 100 AI questions per month, and 1 seat — no credit card required.
How much is Pro?
$99/month billed monthly, or $79/month billed yearly. Includes 5 knowledge bases, 1,000 AI questions per month, 10 seats, Slack / Discord / Teams bots, Support Ticket AI, MCP server, and analytics.
What counts as an AI question?
Each end-user message the assistant answers counts as one question — whether it comes through the widget, a bot (Slack / Discord / Teams), the browser extension, Support Ticket AI, MCP, or the public API. Follow-up messages in the same conversation each count as a question. Search API queries (retrieval without generation) do not count.
Can I try Pro before paying?
Yes. Pro comes with a 14-day free trial. No credit card required to start.
What happens if I exceed my monthly question limit?
You're notified at 80% of your quota. Once you hit 100%, new questions are rejected with a specific 'monthly quota exceeded' response — a hard limit, not a suggestion, so you don't get a surprise LLM bill. Upgrading takes effect immediately; otherwise the quota resets on your billing date.
Do actions or MCP calls count against my quota?
Only end-user chat questions count against the AI-question quota. MCP list_sources and search_docs (retrieval-only) don't count; ask_question calls do.
What's on Enterprise?
Custom quotas, SSO (OIDC / SAML), DPA, HIPAA BAA, regional data residency (US / EU), BYOK (bring your own OpenAI / Anthropic / Gemini key), hide-branding on the widget, dedicated support, custom SLA. Contact us for pricing.