Platform · Knowledge base
Agents that answer from your content.
Vocily AI grounds answers in your own knowledge base — websites, documents, FAQs — so the agent stays accurate without you writing scripts for every case.
Problem · Solution
The problem today
An agent without a knowledge base will make things up. Plausible-sounding things. About your product, your pricing, your refund policy. The fix is to write every possible answer into the prompt — but prompts drift, content changes weekly, and the team that knows the policy isn't the team writing the prompt. By the time the agent quotes a wrong refund window on a live call, the trust is gone.
How Vocily AI handles it
Website ingestion
Point Vocily AI at your site. It crawls and ingests pages into a knowledge base — pricing, FAQ, product pages, policies — without re-templating any of it.
Document sources
Add PDFs, internal docs, and policy files alongside web content. Mixed-source KBs are first-class.
Live retrieval, not memorisation
When the agent needs a fact mid-call, it pulls live context from the KB — not from a frozen prompt. Today's content answers today's calls.
Per-agent scoping
Each agent attaches to its own KB. The billing agent doesn't see the sales KB; the sales agent doesn't quote refund policy.
Refresh on demand
When your content changes, re-ingest the source. The agent's answers update with it — no prompt edit required.
Workspace library
Workspaces keep a library of knowledge bases that any agent can reuse — your product KB is built once, attached many times.
What's in it
How knowledge moves into a call.
From source ingestion through to the agent quoting a fact mid-conversation — each step is explicit.
Sources you can ingest
- Websites
- Point at a domain or specific URL list; Vocily AI crawls pages and stores normalised text.
- PDFs
- Upload product, policy, or FAQ PDFs — text and tables are extracted.
- Plain docs
- Markdown, plain text, structured FAQ exports — anything text-based.
- Mixed
- A single KB can mix sources — the agent retrieves across all of them.
How retrieval runs
Pick deterministic for exact-match Q&A, semantic for conceptual questions.
- Deterministic
- Exact-match Q&A — best for policy docs, pricing FAQs, structured help content.
- Semantic
- Embedding-based retrieval — best for conceptual questions where the caller phrases things differently than your docs do.
- Multilingual KBs
- A KB built from English docs still answers callers in Hindi or Tamil — the agent translates retrieved passages on the way out.
- Match
- Most relevant passages for the question are retrieved.
- Grounded answer
- The agent answers using the retrieved passages — not from memory or the prompt.
- Empty-KB fallback
- If the KB has no answer, the agent says so and follows your escalation rule.
Per-agent scoping
What each agent can and can't see.
- Attached KBs
- Each agent declares which KBs it can read from.
- Cross-talk
- Other agents' KBs aren't visible. Sales agent ≠ billing agent ≠ retention agent.
- Reuse
- Same KB can attach to multiple agents — build the product KB once.
- Workspace library
- All KBs live in a workspace library, attachable to any agent there.
Keeping it fresh
When your content changes, the agent updates with it.
- Refresh
- Re-ingest any source on demand. Next call uses the new content.
- Source priority
- If multiple sources answer the same question, you control which wins.
- Versions
- Re-ingestion replaces the source — old answers don't linger in retrieval.
- Health
- Failed crawls or broken sources surface as a workspace warning, not a silent miss.
How grounding works
Live context during the conversation.
When an agent needs a fact mid-call, it pulls live context from its knowledge base — not from a frozen system prompt.
Knowledge attached
Clinic reception · ag_clinic_v3
Website
clinic.example
Documents
Pricing & services
Documents
Common FAQs
Common questions
What teams ask before they switch.
The agent retrieves the most relevant passages for each question and grounds its answer in them. If the KB doesn't have the answer, the agent says so and escalates per your rules — it does not invent.