Survey answers grouped with Qdrant and summarized by AI

Survey answers grouped with Qdrant and summarized by AI

An n8n workflow that turns a sheet of survey answers into a sheet of insights. It reads every answer from Google Sheets and
stores each one as a vector in Qdrant, with its question and its participant. For each question it sorts the stored answers
into groups of similar answers with the K-means method. An OpenAI model then summarizes each group with an insight and a
sentiment, and every result is added as a row to a new sheet.

What is inside

• workflow.json: the workflow as n8n exports it, 30 nodes and 12 sticky notes. It holds no credential.
• README.md: these notes.

What we changed from the original copy

• Apply K-means Clustering Algorithm was a Python node that imported numpy and sklearn and read its input with
  _input. n8n's code node documentation, read on 12 November 2026, says Pyodide is removed from n8n 2.0. The native
  Python that replaced it offers no _input and cannot import those packages by default. The node is now JavaScript. It
  does the same job: up to 10 groups with a fixed seed, and the same output of an id, a total and the Qdrant points
  of each group.
• Has Clusters? only checked that the list of groups was not empty. K-means always returns groups, so the check passed
  even when every answer was different. Only Clusters With 3+ points then dropped every group. Nothing reached the nodes
  after it, and the run stopped there. No row was written, and the next question never started. The check now counts the
  groups with three or more answers, and a question without one gets its No Insight Found row.
• Find All Answers wrote the question into its JSON text by hand, so a question with a double quote broke the request.
  The body is now built with JSON.stringify.
• Find All Answers and Get Payload of Points called qdrant:6333, the name of a Docker service on their author's
  machine. Both now take the address from the new qdrantUrl value in Set Variables, which holds YOUR_QDRANT_URL.
• OpenAI Chat Model asked for gpt-4o-mini. OpenAI's deprecations page, read on 12 November 2026, lists it for shutdown
  on 14 January 2027, with gpt-6-luna as the replacement. The node is now version 1.2 with that model.
• Get Survey Results pointed at its author's spreadsheet. It is empty now, and you pick yours from the list.
• Export To Sheets carried six sample column names from its author's own run, which included answers from the sample
  survey. They are gone, and the node maps by field name like Export To Sheets1.
• The system prompt of Survey Insights Agent had two typos, "mildy", and they are fixed.
• The 12 sticky notes are rewritten and hold no outside links. The original copy linked to documentation pages, a sample
  sheet and a chat server.
• Every credential reference is removed, so you pick your own.

Before you start

• An n8n instance, on n8n Cloud or your own server.
• A Google account, a spreadsheet of survey answers and a Google Sheets OAuth2 credential. The sheet needs a Participant
  column, and every question column needs a question mark in its header. A column without one is stored in Qdrant too,
  but it is not analyzed.
• A Qdrant server of your own or on Qdrant Cloud, with its address and an API key.
• An OpenAI API key that can use text-embedding-3-small and gpt-6-luna.

Set it up

1. In n8n, create a workflow, open its menu and choose Import from File. Pick workflow.json.
2. Save the workflow. Trigger Insights starts the second part by calling the workflow by its ID.
3. Add a Google Sheets OAuth2 credential. Select it in Get Survey Results, Get Sheet Details, Create Insights Sheet,
   Get Survey Headers, Export To Sheets and Export To Sheets1.
4. In Get Survey Results, pick your spreadsheet and your sheet from the lists. The title of the spreadsheet becomes the
   survey name that ties the answers together.
5. Add a Qdrant credential with the address and the API key. Select it in Qdrant Vector Store, Find All Answers and
   Get Payload of Points.
6. In Set Variables, replace YOUR_QDRANT_URL with the same address, with no slash at the end.
7. Add an OpenAI credential and select it in Embeddings OpenAI and OpenAI Chat Model.
8. Press Test workflow once. The answers are stored first, then the second part starts by itself and writes the insights
   to a new sheet.

How a run goes

1. When clicking ‘Test workflow’ starts the run and Get Survey Results reads every row of the sheet.
2. Convert to Question Answer Pairs turns each row into one pair for every question column, with the participant, and
   QA Pairs to List makes one item of each pair.
3. Qdrant Vector Store stores each answer. Embeddings OpenAI makes its vector, and Default Data Loader attaches the
   question, the participant and the survey name as metadata.
4. Prep Values For Trigger passes on the spreadsheet ID and the sheet, and Trigger Insights starts the second part.
5. Execute Workflow Trigger receives them. Get Sheet Details reads the spreadsheet's title, and Set Variables gathers
   the Qdrant address, the collection name, the survey name and the new sheet's name.
6. Create Insights Sheet adds the sheet. Get Survey Headers reads the first two rows, Extract Questions keeps the
   headers with a question mark, and Questions to List makes one item of each.
7. For Each Question... takes one question at a time. Find All Answers fetches its stored answers with their vectors,
   up to 500.
8. Apply K-means Clustering Algorithm sorts them into up to 10 groups.
9. Has Clusters? sends the run on when some group has three or more answers. Otherwise Prep Values For Export and
   Export To Sheets1 write a No Insight Found row.
10. Clusters To List makes one item of each group, and Only Clusters With 3+ points keeps the big ones.
    Get Payload of Points fetches the answers of each, and Survey Insights Agent writes a summary, an insight and a
    sentiment.
11. Prep Output For Export adds the number of answers, the participant IDs and the raw answers, and Export To Sheets
    adds the row. The loop then moves on to the next question.

Change the look

• The spreadsheet: the document and the sheet in Get Survey Results.
• The Qdrant address: YOUR_QDRANT_URL in Set Variables.
• The collection: collectionName in Set Variables and the collection in Qdrant Vector Store. Keep them the same.
• The name of the new sheet: insightsSheetName in Set Variables.
• The number of groups: the 10 in Apply K-means Clustering Algorithm.
• The smallest group: the test for three or more answers in Has Clusters? and in Only Clusters With 3+ points. Change both.
• What the model writes: the system prompt and the example in Survey Insights Agent.
• The model: OpenAI Chat Model.

Good to know

• Run it once for each survey. A second run stores the same answers again, so groups would count each answer twice. Clear
  the collection in Qdrant first if you must run it again.
• It needs many answers. With about ten answers or fewer, almost every group holds one answer, and the question gets a No
  Insight Found row. A few dozen answers to a question is a fair start.
• Find All Answers fetches at most 500 answers for a question, and nothing pages on after that.
• Qdrant Vector Store creates the collection on its first insert, so you do not make it by hand.
• The embeddings node names text-embedding-3-small, which is not on OpenAI's deprecations page.
• Your data: every answer goes to OpenAI to make its vector, and the answers of each big group go to OpenAI again for the
  summary. The answers and vectors are kept in your Qdrant. The insights and the answers behind them are
  written to your sheet. If your survey holds personal data, keep that in mind.
• The summaries come from a model. Read a few raw answers of each group before you rely on one.