Knowledge-Driven Answers

When a customer asks something the agent doesn’t know off the top of the prompt — a policy detail, a procedure, a step-by-step from your SOPs — Omniflow retrieves the relevant chunks from your internal wiki and grounds the answer in them. This is what makes the difference between an agent that bluffs and an agent that’s right.

The capability in one paragraph

The chatbot is connected to your internal knowledge — Confluence, SharePoint, Notion, Google Drive, your help center, or articles authored inside Omniflow. At runtime, every customer turn triggers a retrieval step: Omniflow embeds the question, queries the agent’s knowledge index, and injects the most relevant chunks into the model’s context. The model answers from those chunks, in your house style, with traceable citations back to the source document.

How it works end-to-end


   1. Customer:  "What's the policy on a refund after 30 days?"
                       │
                       ▼
   2. Embed turn + filter by agent.knowledge_scope
                       │
                       ▼
   3. Vector search → top 5 chunks
                       │
                       ▼
   4. Inject chunks + system prompt + history → LLM
                       │
                       ▼
   5. Agent: "Refunds after the 30-day window can still be issued
              for gold-tier customers per section 4.2 of the
              refund policy. I can process up to $500; above
              that requires manager approval."
                       │
                       ▼
   6. Trace stores: which chunks were retrieved, which were cited.

Why the answer is reliable

Property	How Omniflow guarantees it
Grounded	Retrieval-Augmented Generation (RAG). The model sees your documents, not just its pretraining.
Cited	Every retrieval is logged in the trace with the source URL and chunk text.
Refusal-aware	Prompts include “if you don’t know, say so” guardrails. The agent escalates instead of guessing — see Escalation & Handoff.
Fresh	Source content re-syncs hourly (Confluence, SharePoint, Notion) or on save (Omniflow KB). Stale answers don’t linger.
Permission-aware	The integration user’s read permissions are honored. The agent can’t quote a wiki page it doesn’t have access to.

Why the answer is structured

The model is prompted to return answers in a consistent shape per channel:

Voice — short conversational reply, then offers to send detail by email or SMS.
Chat — bullet list or numbered steps when the answer has structure; plain paragraph otherwise.
Email — full structured response with optional attachments.

Tone and structure are tunable in Prompts & Tone.

Why the answer is contextualized

Every reply has access to:

The full conversation history — including across channels for authenticated users (chat history visible to a voice agent on the next call).
The customer’s CRM record — tier, account owner, lifetime value, custom fields. See CRM integrations.
The agent’s tools — booking lookup, order status, account state. See Tools & Actions.
Time and locale — business hours, language, region.

The retrieved knowledge isn’t applied in a vacuum — it’s combined with everything Omniflow knows about the customer.

Set it up

Connect your wiki

Pick the source — Confluence, SharePoint, Notion, Google Drive, a public help center URL, or articles inside the Omniflow KB. See Knowledge integrations.

Scope the agent’s knowledge

Open the agent → Knowledge tab → pick spaces, libraries, or labels the agent can retrieve from. Different agents can have different scopes (a customer agent vs an internal IT agent).

Pick a retrieval mode

Always retrieve (default) — every turn does a retrieval. Best for support agents.
On-demand — the model calls a search_knowledge tool when it decides retrieval is needed. Cheaper for high-volume agents with lots of small talk.
Off — pure prompt mode. Only useful for very narrow, scripted agents.

Tune for reliability

Top-K — how many chunks to inject. Default 5.
Min-score — drop chunks below a confidence threshold so the model isn’t fed weak matches.
Refusal threshold — if the best chunk’s score is below this, the agent says it doesn’t know and offers to escalate.

Test

In the agent sandbox, ask a question whose answer is in your wiki. Verify the trace shows the right chunk was retrieved. Iterate on the article or the chunking if the wrong chunk wins.

Reliability guardrails

Hallucination is a real risk for any LLM. Omniflow mitigates it with retrieval, refusal prompts, and citation traces — but the right defense is a tight rubric in QA so you catch any hallucination as soon as it ships, and feed the failure back into knowledge improvement.

No-answer fallback. When retrieval can’t find a confident chunk, the agent says it doesn’t know and offers a handoff path.
PII redaction. PII is masked on the way into traces and logs, configurable per workspace.
Source verification. Every cited source is link-checkable from the trace. If the source page is gone, the trace flags it on the next sync.

Compliance considerations

Concern	Treatment
Sensitive content in the wiki	Don’t sync spaces the agent shouldn’t quote. Scope tightly.
Audit trail	Every retrieval is logged immutably in the trace. Export via API.
Right to know what data was used	The full retrieval list is in the trace, including chunk text.
Data residency	Knowledge embeddings stay in your workspace’s region; never cross to another region.

Open in Omniflow

If you want to…	Go to
Connect Confluence or SharePoint	Knowledge integrations
Tune the agent’s knowledge	Knowledge Sources
See what was retrieved on a call	Activity Logs & Traces
Hand off when the agent doesn’t know	Escalation & Handoff