Study notes from files dropped in a folder

Study notes from files dropped in a folder

An n8n workflow that watches a folder and writes study notes for every new file. Drop in a PDF or a text file. The workflow stores it
in Qdrant, asks Mistral questions about it and writes three Markdown documents: a study guide, a timeline and a briefing.

What is inside

• workflow.json: the workflow as n8n exports it, 40 nodes, seven of them sticky notes. It holds no credential. Five notes are
  coloured frames, one holds the try-it summary, and one is a small hint about adding templates.
• README.md: these notes.

What we changed from the original copy

• The Mistral and Qdrant nodes were linked to credentials in their author's n8n. The links are gone, so select your own.
• All four chat models were open-mixtral-8x7b, which Mistral retired on 30 March 2025, or had no model and used n8n's old default,
  mistral-small. They are now set to mistral-small-latest. Both embeddings nodes are set to mistral-embed, which they used
  before without saying so.
• The notes were written into the watched folder. Each new note then started the workflow again, and every run wrote more notes.
  They now go to a folder named notes beside the watched one, and the sticky note about the export says so.
• The Local File Trigger now waits until a file has finished copying before it starts.
• To Binary named its source field with an expression that returned the whole text, so it looked for a field with that text as its
  name and failed. It now names the field data.
• Text and Markdown files were read into a field called data, but the next step reads text. Extract from TEXT now writes
  text, as the PDF step does.
• The Word branch is gone. A file type of docx never occurs, so the branch was never reached, and its node was set to the
  OpenDocument spreadsheet operation. Any file that is not a PDF goes down the text path.
• Interview returned one item per question, but a Split Out step after it looked for a list field and found none. The step is
  gone, and Answer from Qdrant reads each question from the text field.
• The Qdrant collection is chosen by name, so it need not exist yet. 2secs is renamed Pause two seconds and its amount is set to
  two seconds. Before, it used n8n's default.
• Nothing else.

Before you start

• A self-hosted n8n. The Local File Trigger does not exist on n8n Cloud.
• Access to a folder that n8n can read and write. In Docker, mount a host folder at /home/node/storynotes.
• A Qdrant server, and its address and API key if it needs one.
• A Mistral API key.

Set it up

1. Turn on the Local File Trigger. n8n 2.0 and later leave it off. Set the environment variable NODES_EXCLUDE to [] and restart
   n8n.
2. Allow n8n to use the folder. Set N8N_RESTRICT_FILE_ACCESS_TO to /home/node/storynotes, since n8n 2.0 limits file access to
   one folder otherwise. Create two folders inside it, context and notes.
3. In n8n, create a workflow, open its menu and choose Import from File. Pick workflow.json.
4. Add a Qdrant credential with your server's address and key. Select it in Qdrant Vector Store and Qdrant Vector Store1.
5. Add a Mistral Cloud credential. Select it in the four chat models, Mistral Cloud Chat Model through Mistral Cloud Chat Model3,
   and in Embeddings Mistral Cloud and Embeddings Mistral Cloud1.
6. Publish the workflow. Copy a PDF or a text file into the context folder.
7. Wait a few minutes, then look in the notes folder for three files, one for each template.

How a run goes

1. Local File Trigger fires when a new file appears in /home/node/storynotes/context. Settings takes the project name from
   the fourth part of the path, storynotes, and keeps the path and the file name.
2. Import File reads the file and Get FileType sends a PDF to Extract from PDF and everything else to Extract from TEXT.
   Prep Incoming Doc puts the text in a field called data.
3. Two things then happen at once. Qdrant Vector Store stores the text, cut into pieces of 2,000 characters by `Recursive Character
   Text Splitter and embedded by Embeddings Mistral Cloud, with the project and file name as metadata. Summarization Chain`
   writes a summary in chunks of 4,000 characters.
4. Merge waits for both. Prep For AI gathers the summary, the project and the file name, and Get Doc Types lists three
   templates: a study guide, a timeline and a briefing doc. Split Out Doc Types makes one item per template.
5. For Each Doc Type... takes one template at a time. Interview reads the summary and writes questions that the template would
   need answered.
6. Answer from Qdrant answers each question. Vector Store Retriever fetches the closest four pieces, using
   Qdrant Vector Store1 and Embeddings Mistral Cloud1. Aggregate collects the answers.
7. Generate writes the document from the template's description and the answers. Pause two seconds waits and the loop moves
   to the next template.
8. When all three are done, Get Generated Documents, To Binary and Export to Folder write each one as a Markdown file in
   /home/node/storynotes/notes. The name is the first 21 characters of the source file name, three dots and the template title.

Change the look

• Which templates are written: the list in Get Doc Types. Add an entry with a file name, a title and a description.
• How many questions are asked: the wording in Interview.
• Where the notes go: the file name in Export to Folder.
• How a document is written: the instruction in Generate.
• The models: the model field of each chat node, and of the two embeddings nodes. Both embeddings nodes must use the same model.
• How big a stored piece is: the chunk size in Recursive Character Text Splitter.

Good to know

• Self-hosted only: n8n's docs say the Local File Trigger is not on n8n Cloud, and from n8n 2.0 it is off until you change
  NODES_EXCLUDE. Anyone who can edit workflows on your n8n can then watch folders on the machine, so enable it only where you
  trust them.
• Word files: the workflow reads PDFs and plain text. n8n's Extract From File node has no Word operation, so a .docx takes the
  text path and comes out as noise. Save it as PDF first.
• Models: Mistral's models page, read on 12 November 2026, lists open-mixtral-8x7b as retired on 30 March 2025, with Mistral Small 4
  as the alternative. The card for Mistral Small 4 names mistral-small-latest as one of its names. The chains here use no tools.
• One collection: every file goes into the same Qdrant collection, storynotes, and Answer from Qdrant searches all of it. Questions
  about one file can pull in pieces of another. The metadata holds the file name, so a search filter on it would keep them apart.
• Cost: each file means about 21 model calls for the three documents, plus the summary calls and one embedding call for each piece.
• Your data: the text of every file goes to Mistral and to Qdrant.
• Not run: nothing here was run against Mistral or Qdrant, and no file was watched. The edits come from n8n's source and the
  vendors' pages, read on 12 November 2026. Run one short file first and read the notes before you trust it.