Cited sources
Every AI answer points back to the source document, paragraph, or FAQ entry. Visitors can read the original. Auditors can review the chain.
Feature · Knowledge
Knowledge base AI that cites its sources
OrcaLinq grounds AI answers in your real business knowledge — and pages a person when the knowledge is thin or the topic is sensitive.
What it is
RAG that knows when not to answer
Retrieval-augmented generation grounded in your real business knowledge. Every answer cites its source. When confidence is thin, the AI asks for a person instead of guessing.
Confidence-aware retrieval
If retrieval scores are low or the question doesn't match indexed knowledge, the AI doesn't fabricate — it offers a person.
Per-tenant isolation
Knowledge, embeddings, and indexes live in your tenant only. They are not shared with other tenants or used to train shared models.
Closed-loop gap detection
When the AI hits a question it can't answer well, the gap is logged. Owners see the gap on the Podium and can add the missing answer in one click.
What goes in
Knowledge that earns its place
The platform supports both structured FAQ entries and full document ingestion. Pick the format that matches the content.
- FAQ entries. Question + answer pairs. Best for canonical, short policies.
- Documents. PDFs, Markdown, HTML exports from a CMS. Chunked, embedded, cited.
- Web crawl. Public website pages, ingested with respect for robots.txt.
- Knowledge from conversations. Approved answers from real human handoff become future AI answers.
- Live tools. Calendar, CRM, order lookup, inventory check — connected, not memorized.
Safety
What we do to keep AI honest
RAG is not a free pass. The platform layers conservative defaults on top of retrieval to limit hallucination and respect tenant policy.
Refusal over guessing
If retrieval can't substantiate the answer, the AI says it doesn't know and pages a person. No invention.
Visible citations
Cited sources are linked in the response so customers can verify what they were told.
Risk-sensitive routing
Legal, medical, billing, and complaint intents always force a handoff. The AI doesn't speak for the business on sensitive topics.
Per-tenant model keys
Bring your own AI key on Agency to keep request and response data inside your provider account.
Knowledge FAQ
Frequently asked questions
What goes into the knowledge base?
FAQ entries, policy documents, product specs, pricing details, return rules, shipping windows, opening hours, escalation contacts. Anything that staff currently answer manually.
Can the AI hallucinate?
It can — every LLM can. The product is engineered to minimize this with confidence scoring, source citation, and explicit handoff when knowledge is thin. Risk-sensitive intents always route to a person.
Does it work with private knowledge?
Yes. Knowledge is per-tenant. Documents and embeddings are scoped to one tenant and never shared with another tenant or with the model provider's training pipeline.
What about live data like orders or inventory?
Live systems are connected as tools, not as static knowledge. The AI calls the tool when it needs real-time data; it doesn't try to memorize prices or stock levels.
How are knowledge gaps handled?
When the AI hits a question it can't answer confidently, it logs a knowledge gap, hands off to a human, and surfaces the gap on the Podium so owners can add the missing answer.
Related pages
Keep exploring
See your knowledge answer real customer messages.
Bring an FAQ document or one of your support macros. We'll ingest it during the demo and run real questions through it.