An answer with nothing behind it
A chatbot with no "I don't know" will answer from general knowledge even when your documents say something different, or say nothing at all — and the reader cannot tell the difference.
A knowledge base you can type to, or talk to
It doesn't answer
unless it can point to where.
Every answer carries its source, or it is not given at all. When nothing in the index covers the question, it says so and hands off to a human — and that handoff is confirmed by reading the reply back, never assumed.
Source or silence
No sourced answer means an honest handoff, not a guess
Two ways in
Type it or talk to it — the same index answers either way
Confirmed handoff
Escalation is verified by reading the reply back, never assumed
Loaded on click
The voice agent costs nothing until you actually ask it something
Try it
Ask it something, on a public demo corpus
Type to it or talk to it — same index, same sourcing rule either way. This is the real agent, not a recording.
Ask about the F1 2026 rulebook, the Philippine Constitution, the 5 Types of
Wealth, the Almanack of Naval Ravikant, or John's services.
Loading the voice agent…
The blind spot
A confident wrong answer is worse than no answer
Most chatbots are built to always sound sure. A wrong answer stated plainly costs more than an honest "I don't have that," because the reader has no reason to doubt it.
An escalation nobody confirmed
"Passed to a human" is easy to say and easy to get wrong. If the handoff itself fails silently, the question is unanswered and nothing anywhere says so.
A safety net nobody has watched catch anything
A fallback path that has never actually been exercised is a claim, not a proof. The only way to know it works is to break it on purpose and watch what happens.
How it works
One index, two front doors, one confirmed exit
Typing and talking are two ways into the same answer, not two separate products. Both end the same way when the index has nothing: a human, and proof the human was actually reached.
-
01
Index the documents
Source documents are embedded once and stored in a vector index. Nothing is answered from outside that set.
PineconeAzure OpenAI embeddings -
02
Answer, with a source
A typed or spoken question is matched against the index. A reply is only given alongside the document and location it came from.
n8nLangChain -
03
Or refuse, out loud
When nothing in the index covers the question, the agent says so in plain language rather than answering from general knowledge.
FastAPI -
04
Hand off, and confirm it landed
A refusal escalates to a human over Telegram. The handoff is checked by reading the delivery reply back, not assumed from a queued request.
Telegram Bot APIpytest
The evidence
Three real exchanges, one of them a deliberate failure
Captured from the acceptance runs on 2026-08-29. Recorded, not staged for this page — including the one where the safety net was cut on purpose.
Q: what is the minimum mass of the car?
During Sprint Qualifying and Qualifying the Minimum Mass is 726kg plus the Nominal Tyre Mass. In all other sessions it is 724kg plus the Nominal Tyre Mass.
Answered from a source document, and it names the source out loud.
Q: what is the capital of Peru?
I don't have a sourced answer for that, so I'm passing it to a human who will follow up.
Escalation delivered — confirmed by message id
Nothing in the knowledge base covers this, so it refused rather than guessed — and the handoff was confirmed by reading the reply back, not assumed.
Q: how do I cancel my gym membership?
I don't have a sourced answer for that, so I'm passing it to a human who will follow up.
Escalation did not reach a human — this question is unanswered. (Unauthorized: invalid token specified)
The channel was deliberately broken. The agent still spoke truthfully, but the handoff never arrived — and the console said so instead of staying quiet. A safety net nobody has seen catch anything is not known to work.
What you get
Four things, in plain language
Written to be read by whoever is deciding whether to trust it with real documents.
Sourced answer
Every reply names the document and location it came from, so it can be checked without asking the agent to explain itself.
Answers: where did this come from?
Confirmed escalation
A refused question hands off to a human, and the handoff itself is verified rather than assumed to have arrived.
Answers: did the escalation actually land?
Two ways to ask
Type it or talk to it — same index, same sourcing rule, same refusal behaviour either way.
Answers: does it work the way people actually ask?
A demo corpus, named as one
The public index is a demo document set, not a client's real corpus — the difference matters, so it's stated rather than implied.
Answers: is this open to strangers, or a demo?
Try it on the live site, or ask about your own documents
The typed and spoken versions above are running today against a public demo corpus. Wiring the same agent to a real document set is a separate, scoped conversation.