Konkui AI

Thai online sellers get customer questions on LINE, Facebook and Instagram at once, and most of them repeat — price, stock, shipping, store policy — answered by an admin one message at a time. Handing that to an AI is risky: a wrong price or stock level does more damage than no answer at all. The brief: an assistant that answers from the shop’s own data from day one, never states a number it didn’t look up, and hands the chat to a person the moment it can’t help.

The hard part of a sales chatbot isn’t answering — it’s not answering wrong. So a second model reviews every reply before it goes out, and the data the shop enters is the only source of truth.

1. 1. Multi-tenant from day one

   Each shop is a Workspace whose data is fully isolated, and every query is scoped to it. Members are Admins or Operators and sign in with Google. Built on Next.js 16 with Postgres, pgvector and Drizzle.

2. 2. Channel adapters that follow each platform’s rules

   A LINE OA connects with its Channel ID, secret and access token, and only turns Connected once a signed webhook ping is verified; Facebook Pages and Instagram connect through Facebook Login. On LINE the Assistant uses the free reply window before falling back to a push, and on Meta it only sends inside the 24-hour window.

3. 3. Data the Assistant can trust

   A shop keeps Products, Promotions, Bookings and Customers. Core fields carry a role — price, stock — that tells the Assistant what they mean, and Promotions apply automatically inside their date window. The AI can write Bookings and Customers but never search them, so one customer can never be shown another’s details. All seven source types are ingested in the background, and instruction-like text is flagged on the entry.

4. 4. A judge before every action

   The Assistant retrieves Knowledge by meaning, in Thai and English, through embeddings. Before any write and before a reply is delivered, a judge model reviews it: a reply quoting money or stock that no tool returned is rejected and gets one repair, and if the judge can’t run, the chat is handed to the team.

5. 5. A learning loop people approve

   A thumbs-down with a correction enters a queue that an Admin approves into a Q&A pair. The Curator turns context an Admin pastes into proposals to review, and outside AIs such as Claude Code, Codex or Cursor can propose Knowledge over MCP — every proposal waits for a person. The Playground shows each reply’s match scores, tool calls, judge verdict and cost.

Konkui AI runs in production. LINE OA connects today; Facebook Pages and Instagram accounts belonging to other businesses show ‘Review pending’ until Meta’s App Review is complete.

• Every chat in one Inbox

  The team sees every conversation with who holds it — Assistant, Operator, waiting or resolved — and each shop’s data stays in its own Workspace.

• Answers grounded in the shop’s data

  The Assistant answers product, price and promotion questions from what the shop entered, and says so plainly when a fact is missing instead of guessing.

• A person takes over cleanly

  On handoff the Assistant stops replying, but drafts a suggested reply in the composer that the Operator can use, edit or discard.

• Corrections become Knowledge

  Once approved, a team correction is used by the Assistant from the very next message.