// Module 02 — Knowledge Engine
Correct an answer once — and every channel speaks the new truth the same turn. Fix an answer once — every channel says the new, correct thing right away.
The Knowledge Engine is the brain behind every reply your platform gives. It grounds the chatbot and the phone agent in one shared, self-correcting source of truth — so when your team fixes something, the fix is live in the next message, on every channel, with no retraining and no developer.
The Knowledge Engine is the brain behind every answer your platform gives. It keeps the chatbot and the phone agent reading from one shared set of facts that corrects itself — so when your team fixes something, the fix shows up in the very next message, on every channel, with no retraining and no developer needed.
// Instant live corrections
Your team's note is the supreme source of truth — and it lands the next turn Your team's note is the final word — and it takes effect right away
Operators correct any answer straight from the Command Center, and the fix is live in the very next message — across chat and phone — with no retraining, no developer, and no redeploy. Standing operator notes are the highest authority in the system: they override stale training data and built-in facts the moment they're written. If yesterday's hours were wrong, you type the right ones once and both the web assistant and the 24/7 phone line start saying them immediately.
Your team can fix any answer right from the Command Center dashboard, and the fix shows up in the very next message — on chat and on the phone — with no retraining, no developer, and no waiting for a new release. A note your team writes and keeps in place is the highest authority in the system: it overrides old training data and built-in facts the moment it's written. So if yesterday's hours were wrong, you type the right ones once, and both the web assistant and the 24/7 phone line start saying them right away.
Standing notes are the supreme source of truth. They override stale data within a fixed safety floor — so a correction takes effect instantly, but it can never push the assistant past the rules it isn't allowed to break.
A note your team keeps in place is the final word. It overrides outdated information, but only within a fixed set of safety rules — so a correction takes effect right away, but it can never push the assistant past the rules it's not allowed to break.
- Fix any answer from the dashboard — the next chat and phone reply uses it.Fix any answer from the dashboard — the very next chat and phone reply uses the fix.
- No retraining, no model fine-tuning, no developer, no redeploy.
- No retraining, no model adjustments, no developer, no new release needed.
- Standing notes outrank stale training data and built-in facts.
- A note your team keeps in place outranks old training data and built-in facts.
- One correction propagates to every channel at once, the same turn.
- One fix reaches every channel at once, right away.
// Stays current automatically
The knowledge base keeps itself fresh while you run the day The knowledge base stays up to date on its own, all day
Truth drifts when nobody's watching it. The Knowledge Engine refreshes from your own website every day, so the things you publish stay in sync with what the assistant says. And the email your team already sends becomes knowledge too: staff blasts can be promoted into the knowledge base — with a human review gate standing between an inbound message and what your assistant tells the public.
Information goes stale when no one is watching it. The Knowledge Engine checks your own website every day, so what you publish stays in sync with what the assistant says. And the emails your team already sends become knowledge too: staff blasts can be added to the knowledge base — with a person reviewing them first, before any message from the outside reaches what your assistant tells the public.
- Daily self-refresh from your own site — published changes flow into the knowledge base without anyone re-typing them.
- Updates every day from your own site — changes you publish flow into the knowledge base without anyone retyping them.
- Email-to-knowledge: marketing-style staff blasts can be promoted into knowledge, so the assistant answers the way your team actually writes.
- Email becomes knowledge: staff blast emails can be added to the knowledge base, so the assistant answers the way your team actually writes.
- Human review gate: threaded conversations and anything ambiguous wait in a review queue for a person to approve before it ever goes live.
- A person checks it first: conversation threads and anything unclear wait in a review queue for a person to approve before it ever goes live.
- Internal stays internal: notes marked internal or do-not-touch are excluded end to end — they never reach a public answer.
- Private stays private: notes marked internal or do-not-touch are kept out of the system completely — they never end up in a public answer.
// Grounded, not invented
It answers from your knowledge — or it declines. It never invents. It answers from your real information — or it says it doesn't know. It never makes things up.
The Knowledge Engine is built accuracy-first: a no-fabrication rule enforced in code, not just suggested in a prompt. On the questions where a wrong answer does real damage — your hours, donation details, endorsements, who you are and aren't affiliated with — cite-or-decline guards require the answer to be grounded in your knowledge. If it isn't, the assistant declines or escalates to your team rather than guessing. That's the floor an operator note can speed up, but cannot switch off.
The Knowledge Engine puts accuracy first: it's not allowed to make things up, and that rule is built into the code itself, not just a suggestion in a prompt. On the questions where a wrong answer causes real harm — your hours, donation details, endorsements, who you are and aren't connected to — the system requires the answer to come from your own knowledge. If it can't find a real answer, it says so or hands the question to your team instead of guessing. Your team's notes can speed this up, but they can never turn this rule off.
This is the safety floor a note can't override: never fabricate hours, dollar figures, endorsements, or names; ground the answer or escalate. You can correct what the assistant says — you can't talk it into making something up.
This is the safety rule a note can never override: never make up hours, dollar amounts, endorsements, or names; answer from real information or hand it to a person. You can correct what the assistant says — you can't get it to invent something.
- Cite-or-decline grounding on the high-stakes facts — hours, donations, endorsements.
- On the facts that matter most — hours, donations, endorsements — it answers from real information or says it doesn't know.
- Uncertain or sensitive questions escalate to your team instead of being guessed at.
- Unclear or sensitive questions go to your team instead of being guessed at.
- The no-fabrication rule lives in code, beneath every operator setting.
- The no-making-things-up rule is built into the code, underneath every setting your team controls.
- A correction makes the assistant faster and sharper — never looser.
- A fix makes the assistant quicker and more accurate — it never makes the rules looser.
// How retrieval works
How it finds the right answer, in plain English
You don't need to think about any of this to run it — but here's what's happening under the hood, without the jargon. When someone asks a question, the engine first does a fast keyword-style search across your knowledge to pull the passages most likely to hold the answer. It can then re-rank those candidates by meaning, so a question phrased differently than your notes still finds the right one. The heavy lifting is done ahead of time, so there's no waiting penalty added to each message — answers stay quick whether it's a slow afternoon or a flood of questions at once.
You don't need to understand any of this to use it — but here's what's happening behind the scenes. When someone asks a question, the system first searches your knowledge for the keywords most likely to hold the answer. It can then double-check that list by meaning, so a question asked in different words than your notes still finds the right one. All the hard work happens ahead of time, so it doesn't slow down any single reply — answers stay quick whether it's a quiet afternoon or a rush of questions at once.
- 1Find the candidates. A fast keyword-style search pulls the passages from your knowledge most likely to contain the answer.Find likely matches. A fast keyword search pulls the passages from your knowledge most likely to contain the answer.
- 2Rank by meaning. An optional meaning-based re-rank reorders those candidates so a question asked in different words still lands on the right note.Double-check by meaning. An optional step reorders those matches by meaning, so a question asked in different words still lands on the right note.
- 3Ground the reply. The assistant answers from what it found — or declines and escalates if it can't ground the answer.Answer from real information. The assistant answers from what it found — or says it doesn't know and hands it to your team if it can't.
- 4No per-message penalty. The preparation is done in advance, so retrieval doesn't slow down individual replies, even under a surge.No slowdown per message. The prep work is done ahead of time, so looking things up doesn't slow down individual replies, even when a lot of questions come in at once.
The takeaway for you: the assistant pulls the right facts quickly, phrases its reply naturally, and stays fast — and every word is grounded in the knowledge you control.
The takeaway for you: the assistant finds the right facts quickly, writes its reply naturally, and stays fast — and every word comes from the knowledge you control.
// Facts graph
See how your key facts connect — and where each one came from. See how your key facts connect to each other — and where each one came from.
The facts graph is an operator-facing audit view, built deterministically from the Campaign Facts you curate in the Command Center — your organization, its endorsements, contacts, and key resources. It lays out how those facts relate and, for each one, shows its source: a synced note, a panel edit, or the shipped default. At a glance you can confirm exactly what the assistant will hand a supporter, and why.
The facts graph is a view for your team, built directly from the facts you manage in the Command Center — your organization, its endorsements, contacts, and key resources. It shows how those facts relate to each other, and for each one, where it came from: a synced note, a dashboard edit, or the built-in default. At a glance, you can confirm exactly what the assistant will tell someone, and why.
- Deterministic, not guessed. The graph is derived purely from the facts you've entered — no LLM, no name-guessing — so it never invents an entity or a relationship you didn't put there.
- Built from your facts, not guessed. The graph comes purely from the facts you've entered — no AI guesswork — so it never invents a person, place, or connection you didn't put there.
- Relationships, mapped. Endorsements, contacts, and resources are linked to your organization, so the connections between your facts are visible in one place.
- Connections, mapped out. Endorsements, contacts, and resources are linked to your organization, so you can see how your facts connect in one place.
- Provenance on every fact. Each value carries its origin — note, panel, or default — making it obvious which facts a note is currently overriding. Notes are the supreme source of truth, and the graph shows it.
- Every fact shows where it came from. Each value carries its source — note, dashboard, or default — so it's obvious which facts a note is currently overriding. Notes are the final word, and the graph shows that.
- Live from your facts. The view reflects your current Campaign Facts the moment you open it — there's no separate build to run and no schedule to manage.
- Always current. The view shows your current facts the moment you open it — there's nothing separate to run and no schedule to manage.
// Live demo
See it answering real questions right now Watch it answer real questions, right now
The live demo runs the exact system that ships to customers — try it as a campaign or as a business.
This demo is the exact same system real customers get — nothing watered down. Try it as a campaign or as a business.