Health plan quote chatbot with Postgres memory and tools

Health plan quote chatbot with Postgres memory and tools

An n8n chat in Portuguese for comparing health insurance plans for companies. An AI agent remembers each visitor in a
Postgres table. It can search a MySQL table of plans for the three cheapest matches by city, number of lives and the holder's
age, and it can call two web APIs. It is a starter: the tables and the two APIs are yours to supply, and the two system
messages are short drafts to rewrite.

What is inside

• workflow.json: the workflow as n8n exports it, thirteen nodes and one sticky note. It holds no credential.
• README.md: these notes.

What we changed from the original copy

• The two steps that called an OpenAI Assistant used n8n's Assistant node. OpenAI's deprecations page, read on 12 November 2026,
  gives 26 August 2026 as the shutdown of the Assistants API, with the Responses API and Conversations API as the replacement.
  Both are now AI Agents, Save Lead Profile and Quote Assistant, each with a model node on gpt-5.6-terra:
  Profile Chat Model and Assistant Chat Model.
• The author's assistants held their own instructions, which are not in the file. The two system messages are ours. A short one
  saves the visitor's profile and confirms in one line. The other tells the quote agent to answer in Portuguese, use its three
  tools and quote only what a tool returns. Read both and rewrite them for your business.
• The tool Products in Daatabase is now Products in Database, and its description was written for it.
• The description of Knowledge Base was filler text. It now says what the tool looks up. Its address pointed at the
  author's company, and it is now https://<YOUR_QUOTE_API_HOST>/widget/info.
• External API called a private host. It is now https://<YOUR_LOOKUP_API_HOST>/findByNameAndBirthDate. Its body wrote
  {{json.name}}, which n8n reads as an expression and cannot resolve. It now uses {name} and {birthdate}, the single-brace
  placeholders an HTTP Request tool fills.
• The chat's greeting named the author's company and a bot. It is now a neutral Portuguese greeting.
• The OpenAI, MySQL and Postgres nodes lost their links to credentials in the author's n8n. Select your own.
• Nothing else. The file has no Code node.

Before you start

• An n8n instance, on n8n Cloud or your own server.
• An OpenAI API key that can use gpt-5.6-terra.
• A Postgres database for the chat memory. The table aimessages is created on first use when it does not exist.
• A MySQL database with a table Products, and a user that can only read it. The query uses these columns: cityQuery,
  state, modality, removed, minLifeAmount, maxLifeAmount, createdAt and ten price columns from priceAtAge0To18 to
  priceAtAge59To199.
• An address for each of the two web tools, or a decision to delete them.
• Your own web page or app, if you want the visitor's profile saved. See the notes below.

Set it up

1. In n8n, create a workflow, open its menu and choose Import from File. Pick workflow.json.
2. Add an OpenAI credential and select it in Profile Chat Model and Assistant Chat Model.
3. Add a Postgres credential and select it in Postgres Chat Memory and Postgres Chat Memory1.
4. Add a MySQL credential for a read-only user and select it in Products in Database.
5. In Knowledge Base, replace <YOUR_QUOTE_API_HOST> with your quote service. In External API, replace
   <YOUR_LOOKUP_API_HOST> with your lookup service. If you have neither, delete the tool and its line in the system message of
   Quote Assistant.
6. Read the system messages of Save Lead Profile and Quote Assistant and rewrite them.
7. Save the workflow, open the chat with the Chat button on the canvas and ask for a plan for a company, with a city, a number
   of lives and the holder's age.

How a run goes

1. Chat Trigger takes a visitor's message. It is public, and it opens with a Portuguese greeting.
2. If checks whether the message carries a leadData object. Without one, the message goes straight to Quote Assistant.
3. With one, Edit Fields1 turns the profile into a sentence: name, age, city, state, profession, education, entity, device,
   channel and plan type. Save Lead Profile stores it in memory and confirms in one line.
4. Edit Fields2 puts the visitor's own message and session_id back, and passes them to Quote Assistant.
5. Quote Assistant answers with Postgres Chat Memory, which keeps the last thirty interactions. It can call
   Products in Database, Knowledge Base and External API.
6. Products in Database returns the three cheapest plans in the modality PME for the city, state, number of lives and the
   holder's age, newest first among equal prices.
7. The reply goes back to the chat.

Change the look

• How the agent talks and what it must do: the system message in Quote Assistant.
• The profile sentence: the text in Edit Fields1.
• The query: Products in Database, including the modality and the number of results.
• The greeting: the initial message in Chat Trigger.
• The tools' addresses and descriptions: Knowledge Base and External API.
• The models: Profile Chat Model and Assistant Chat Model.

Good to know

• The profile branch needs a front end of your own. The built-in chat page sends no leadData and no session_id, so the
  workflow skips the profile step and goes straight to Quote Assistant. A front end can post both fields at the top level of its
  request to the chat's webhook.
• The memory key joins session_id and the chat's own sessionId. Edit Fields2 leaves sessionId out, so a front end that
  sends both would start a new memory on the next message. Send one of the two ids.
• Products in Database writes values the model fills in straight into its SQL text. Give it a user that can only read the
  one table, and think about moving the values into the node's query parameters before real use.
• The chat is public. Anyone who knows its address can talk to the agent and spend your OpenAI key.
• The profile holds a person's name, age, city, profession and education, and it is stored in Postgres and sent to OpenAI.
  External API sends a name and a birth date to your lookup service. Check what your privacy duties allow, for example
  under Brazil's LGPD.
• The query and the tools are tuned to one business: the modality is fixed to PME, and the state in Knowledge Base is fixed
  to SP. Change them for yours.
• The quote agent is told never to invent a plan or a price. Still, test it against your own table before a visitor sees it.
• Each reply can mean several model calls, one for each tool the agent uses.