Deep research report writer with follow-up questions

Deep research report writer with follow-up questions

An n8n workflow that researches a technical question on the web and writes a cited report. It asks follow-up questions
first, searches with Tavily in up to three passes, and turns each report into a Google Doc through a converter you supply.

What is inside

• workflow.json: the workflow as n8n exports it, 90 nodes, no sticky notes. It holds no credential, no key and nothing
  from its author's account.
• Three repairs in the file itself, listed under Good to know: restored backslashes, a placeholder where a converter's
  address stood, and forms without the author's picture.
• README.md: these notes.

Before you start

• An n8n instance, on n8n Cloud or your own server.
• An OpenAI API key. Fifteen nodes use it: fourteen chat models and one call to the Responses API.
• An Anthropic API key for four chat models that repair broken JSON, or OpenAI models in their place (step 3).
• A Tavily API key.
• A Google account for a Google Docs OAuth2 credential, and a converter that builds a Google Doc from Markdown (step 5).

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. Select it in OpenAI Chat Model and OpenAI Chat Model1 to OpenAI Chat Model13, and in
   HTTP Request, the call to api.openai.com.
3. Add an Anthropic credential and select it in Anthropic Chat Model and Anthropic Chat Model1 to
   Anthropic Chat Model3. To skip Anthropic, replace those four with OpenAI chat models: OpenAI Chat Model1 and
   OpenAI Chat Model13 already do that job for two of the repair parsers.
4. Open Tavily Search, Loop2_Tavily Search1 and Loop3_Tavily Search2. In each Authorization header, replace
   <YOUR_TAVILY_KEY> with your key. The brackets go too, so the value reads Bearer, a space, then the key.
5. Open HTTP Request1, HTTP Request2 and HTTP Request3. Replace YOUR_MARKDOWN_TO_DOC_ENDPOINT in each address
   with your converter's address, and select a Google Docs OAuth2 credential (below).
6. Check the model names (chatgpt-4o-latest, claude-3-7-sonnet-20250219, gpt-4.1) against what n8n and your
   accounts offer.
7. Save, open the test address of Start (its path is deep_research) and run one question through it. Then press
   Publish.

The converter it calls

The file does not include it. Each of the three nodes sends two form fields and expects a service that answers by making a
Google Doc:

| Field | What the node sends |
|---|---|
| output | the report in Markdown |
| fileName | the report's title |

The nodes use a Google Docs credential, so n8n adds that credential's access token to the request. Point them at a
service you run yourself. Or delete each of the three and put a Google Docs node set to Create there, with the title
as the document's name, followed by one set to Update that inserts the Markdown as text. The headings then show as
plain text, but no outside service sees your token.

How a run goes

| Stage | Nodes | What happens |
|---|---|---|
| Question | Start, Research Request | A welcome page with a Next button, then a page that asks What would you like to research? |
| Follow-ups | Clarifying Questions to Get Initial Query | The model asks up to three questions, one form page each, or answers No clarification needed. The question and answers become one text. |
| Search phrase | End Form to Prepare Keywords (Code) | The form says thank you and closes. gpt-4.1 through the Responses API, with web search switched on, writes one search phrase. |
| Pass 1 | Tavily Search to Filter Relevant Content per Page | Tavily returns up to ten pages with full text. LLM Relevance Filter keeps up to five. One call reads them and writes an answer with numbered citations. |
| Check | Evaluate Answer Completeness, Check If Query Fully Answered | Answers yes or no. On no, it also writes a new search phrase. |
| Pass 2 | the Loop2_ nodes | The same stage on the new phrase, with its own check, Check Loop2 Answered. |
| Pass 3 | the Loop3_ nodes | The same stage once more. No check follows it. |
| Reports | Prepare Final Report, Final Report Generator and its two copies, HTTP Request1 to HTTP Request3 | Each report is Markdown with a title. The nodes post it to the converter. |

Change the results

• Pages searched and read: max_results and search_depth in the three Tavily nodes (10 and advanced), and the
  up to 5 links line in the three relevance prompts.
• Follow-up questions: the maximum of three in Clarifying Questions and in the description of
  Structured Output Parser1.
• When another pass starts: the criteria in Evaluate Answer Completeness and Loop2_Evaluate Answer Completeness1.
• Report layout: the section list in the prompts of Final Report Generator, Final Report Generator1 and
  Final Report Generator2.
• Models: the model field of each chat model node, and model in Prepare GPT-4.1 Input.
• The kind of question: the examples and the style in Prepare GPT-4.1 Input and the relevance prompts ask for
  tutorials, code and implementation guides. Rewrite them for other research.

Good to know

• The original copy's Code nodes and expressions had lost their backslashes. Twelve nodes read join('n') or
  join('nn') where a line break was meant, and the two Parse LLM Output to JSON nodes had the pattern /{[sS]*}/.
  Mindket put the backslashes back: \n, \n\n and [\s\S]. The twelve are Get Initial Query, the three relevance
  filters, the three page filters, the three report generators, Prepare Final Report and Prepare Final Report Input1.
• The original copy posted its reports to a converter on its author's own domain, with your Google token. This copy has a
  placeholder instead, so no run sends a token anywhere until you choose where.
• The forms showed a banner picture from the author's image host, which this copy leaves out. The welcome text on
  Start said reports were saved to Notion. The file saves through Google Docs, so it now says a Google Doc.
• Passes 1 and 2 each send their answer to a report as soon as it is written, without waiting for the check. A yes from
  a check sends that pass to its report a second time. One run can therefore leave several documents. The pass 2 report
  holds pass 2's answer alone. Only the report from Final Report Generator2 holds all three.
• Done and Done1 are Stop and Error nodes with the message Success, so a run that ends there shows as failed in
  Executions, with that message.
• Prepare Keywords (Code) reads only the first item of the reply's output list. When it finds no search phrase there,
  it falls back to the words fallback search query. If the model runs its own web search first, the message may come
  second, so read the first run's search phrase.
• When the model asks no question, Get Initial Query still writes one stray line into the brief: question: output and
  answer: [object Object].
• Filter Relevant Content per Page and its two copies put up to five full pages into one prompt. One long page can
  run past the model's input limit and stop that step.
• A run makes roughly six to thirteen model calls, depending on how many passes run, plus up to three Tavily searches
  at advanced depth with full page text. OpenAI, Tavily and, for repairs only, Anthropic each bill their part.
  chatgpt-4o-latest is an alias OpenAI moves, and the Responses API call sets store to true.
• Once the workflow is published, anyone with the form's address can start a run on your keys. The form trigger can ask
  for a username and password.
• The question and your answers go to OpenAI. Page text goes to OpenAI and, when a repair is needed, to Anthropic.
  Search phrases go to Tavily, and each report goes to your converter.