AI answers

Two searches run. The passage that wins both writes the answer.

Retrieval runs two searches at once — one by meaning, one by exact term — and fuses the results. The answer is written from the passages that come back and cites which entry it used. There is no path from this to an answer your team never wrote.

The question

how long does delivery to pokhara take?

In your knowledge base · 14 entries

By meaning

Vector similarity over 1024-dimension embeddings

  • Delivery and returns · entry 4
  • Delivery and returns · entry 9
  • Shipping zones · entry 2

By exact term

Keyword scoring — the words the customer actually used

  • Shipping zones · entry 2
  • Delivery and returns · entry 4
  • Trade terms · entry 11

Fused by rank — Ranked highly by both searches, so it survives the fusion.

The answer

Delivery to Pokhara takes 2–3 working days. Order before 2pm and it ships the same day.

Delivery and returns · entry 4

confidence 0.91

Answers

It answers from your documents — and stops when they don't say.

Upload what your team has already written down. Retrieval finds the passage, the model answers from it and cites which entry it used.

Delivery and returns.pdf

In your knowledge base · 14 entries

Orders outside the Kathmandu valley arrive in 2–3 working days. Anything placed before 2pm ships the same day.

Chunked, embedded and searched by meaning and by keyword together. Your documents are never used to train a shared model.

The conversation

Available
how long does delivery to pokhara take?

Answers are retrieved from your own documents and drafted for an agent to send. A newly connected channel starts in suggestions-only mode — the AI drafts, it does not talk to customers until you say so.

States

What the AI is allowed to do, at each stage

A conversation moves through four states, and the state decides what may be sent. This is a field on the conversation, not a mood.

  1. 01

    No AI

    The channel is connected but AI is off. Nothing is drafted.

  2. 02

    AI active

    Answers are found and drafted. On a new connection they are suggestions an agent sends.

  3. 03

    AI unresolved

    No passage covered the question. The thread is flagged and waits for a person.

  4. 04

    Human takeover

    A teammate has the thread. What they answer is kept as a pair, so the next customer to ask is covered.

Boundaries

What it does not do

  • Reply to a customer on a newly connected channel until you turn that on, per connection

  • Answer from anything but the documents you uploaded to that workspace

  • Send your conversation database to the model — the prompt carries the retrieved passages and the message at hand

  • Train a shared model on anything in your workspace

Needs this customer's own order — not in the knowledge base is one of seven recorded handover reasons. Whichever applied is stored on the conversation and shown to whoever picks it up.