Nothing kills trust faster than a phone agent saying you are open until 8 and a chat widget saying you close at 6. Same brand. Same day. Two answers. Caller hangs up. Chat user books the wrong thing. Your front desk spends Monday mopping up fiction.
Spoiler: they do not.
The fix is boring and it works. One knowledge base. Multiple channels. Shared retrieval. Shared update path.
The contradiction problem is a systems problem
People blame “the model.” Usually the model is just reading two different piles of notes.
Voice gets a long system prompt with hours pasted in last spring. Chat gets a Notion export from marketing. Somebody updates the website. Nobody updates either agent. Now you have three truths and a support queue.
If your voice and chat stacks do not share a retrieval source, they will drift. Drift is not a vibe. Drift is wrong answers at scale.
What “one knowledge base” actually means
Do not mean one giant PDF dumped into a vector store and forgotten.
That means:
- canonical docs for hours, locations, programs, pricing rules, FAQs, escalation paths
- versioned content someone owns
- chunked and tagged so retrieval can filter by location or topic
- the same retrieval tool wired into every channel agent
Voice and chat can have different personalities and different turn lengths. They should not have different facts.
Channel differences that are allowed
Voice needs short answers. People cannot hear a paragraph while driving.
Chat can show links, lists, and “tap here to book.”
So keep:
- shared facts in the KB
- channel-specific response style in the system prompt
- shared tools for lookup, ticket create, and handoff
Style is local. Truth is global. That sentence belongs on a sticky note.
How to structure the KB so retrieval does not flop
A few rules I stick to:
One topic per doc when possible. “Facility hours, Downtown” beats a 40-page employee handbook blob.
Explicit effective dates. If insurance rules change in October, say so in the doc title or metadata. Models love outdated certainty.
Escalation snippets as first-class content. Crisis language, after-hours routing, “we cannot answer that” paths should live in the KB, not only in a prompt someone wrote once.
No marketing fluff in ops docs. If the answer is “call admissions,” write that. Do not bury it under brand poetry.
The update workflow that keeps humans honest
The KB needs an owner. Not “the AI team.” A named human who approves changes.
My simple loop:
- Ops or marketing proposes a change in a PR or a reviewed doc edit
- A checklist asks which channels consume this doc
- Change merges
- Agents pick it up on the next sync or immediately if you hot-reload
- Spot-check voice and chat with the same three questions
If updating the website does not update the KB, you still have two brains. Fix the pipeline or accept the contradictions.
Testing for contradictions on purpose
Keep a small eval set: 20 questions that used to cause drama. Hours. Address. Insurance. What to do in a crisis. How to reschedule.
Run them against voice transcript sims and chat side by side. If answers diverge on facts, that is a red build. Style differences are fine. Fact differences are bugs.
This takes less time than one angry customer thread. It also makes prompt thrash less tempting, because the facts are not hiding inside the prompt anymore.
Tooling without turning it into a science fair
You can do this with a managed vector store, a self-hosted DB, or even a well-structured folder of Markdown with good search. The product logo matters less than:
- one write path
- one read API both agents call
- logs that show which chunk was retrieved
If you cannot see what the agent read, you cannot debug why it contradicted itself. Logging retrieval is not optional once money or care is involved.
What to tell clients in one paragraph
Your phone agent and your chat agent are not two products. They are two mouths on one brain. Give them one brain. Make updates painful enough that people do them carefully, and easy enough that people actually do them. Then measure contradictions like you measure downtime.
Bottom line
When that clicks, the weird support tickets drop. The demos stop lying. And your team stops arguing about which prompt is “the real one,” which was never a fun meeting anyway.
