Question answering over local files with Qdrant and Mistral

Question answering over local files with Qdrant and Mistral

An n8n workflow that watches a folder on the machine where n8n runs. It keeps the folder's text files in a Qdrant vector store,
adding new ones, replacing changed ones and clearing deleted ones. A chat then answers questions about the files with a Mistral
model that reads the matching passages. The original notes use bank statements as the example.

What is inside

• workflow.json: the workflow as n8n exports it, twenty-three nodes and six sticky notes. It holds no credential.
• README.md: these notes.

What we changed from the original copy

• The chat model was mistral-small-2402, which Mistral retired on 16 June 2025. The copy calls mistral-small-2603, the
  Mistral Small 4 API name in Mistral's docs, read on 12 November 2026.
• The chat read a collection named BankStatements, while files are stored in local_file_search. Qdrant Vector Store1 now
  reads local_file_search, so the chat finds what the folder branch stores.
• Prepare Embedding Document built each file's location from $json.directory, a field the Read File node does not return. It
  now reads the directory and file name from the file's binary data, so the stored location matches the path the trigger reports.
• Search For Existing Point looked for file_changed on the deleted-file branch. It now looks for file_deleted.
• Both searches asked for one point with limit 1, so only one piece of a file was cleared. The limit is now 1000.
• A sticky note that invited readers to a chat community and a forum lost that last paragraph.
• Credential links and the saved template flag were removed. Select your own credentials.

Before you start

• A self-hosted n8n. The Local File Trigger is not available on n8n Cloud.
• A Mistral API key.
• A Qdrant server that n8n can reach, with its API key if it asks for one.
• A folder of plain text files for the workflow to watch.

Set it up

1. Start n8n with the Local File Trigger allowed. From n8n 2.0 it is excluded by default, so remove it from NODES_EXCLUDE.
2. Let n8n read the folder. From n8n 2.0, N8N_RESTRICT_FILE_ACCESS_TO defaults to ~/.n8n-files, so add your folder to it.
   With Docker, mount a folder at /home/node/host_mount/local_file_search, the path in Local File Trigger, or change the path.
3. Import workflow.json in n8n: create a workflow, open its menu and choose Import from File.
4. Add a Mistral Cloud credential and select it in Embeddings Mistral Cloud, Embeddings Mistral Cloud1 and
   Mistral Cloud Chat Model.
5. Add a Qdrant credential and select it in Qdrant Vector Store, Qdrant Vector Store1 and the four Qdrant HTTP nodes. The
   HTTP nodes also hold the address http://qdrant:6333. Change it if your Qdrant is somewhere else.
6. Activate the workflow. Copy a text file into the folder and watch the run appear under Executions.
7. Open the chat in the editor and ask about the file. The answer should use its contents.

How a run goes

1. Local File Trigger fires for each added, changed or deleted file, after the write has finished. Set Variables turns the
   event into one of file_added, file_changed or file_deleted, with the file's path, and names the collection
   local_file_search.
2. Handle File Event sends the item down one of three outputs.
3. Added: Read File loads the file. Prepare Embedding Document builds one text from the file's location, the processing
   time and the contents. Qdrant Vector Store embeds it with Mistral and stores it. Default Data Loader attaches
   the file's path as filter_by_filename, and Recursive Character Text Splitter cuts the text into chunks. Qdrant creates the
   collection if it is missing.
4. Changed: Search For Existing Point1 asks Qdrant for the points whose filter_by_filename is the path. Has Existing Point?1
   checks there are some and Delete Existing Point deletes them. Remap for File_Added Flow then sends the path down the added
   route, so the file is stored again.
5. Deleted: Search For Existing Point finds the points for the path, Has Existing Point? checks them and
   Delete Existing Point1 removes them. The two delete nodes carry each other's numbers in the original naming.
6. Chat: Chat Trigger sends the question to Question and Answer Chain. Vector Store Retriever fetches the matching chunks
   from Qdrant Vector Store1, and Mistral Cloud Chat Model writes the answer from them.

Change the look

• The folder and the events: Local File Trigger.
• The collection name: qdrant_collection in Set Variables and the collection of Qdrant Vector Store1. Keep them equal.
• How files are cut up: Recursive Character Text Splitter, whose own defaults are 1000 characters and no overlap.
• How many passages the chat reads: Vector Store Retriever, which fetches 4 by default.
• The chat model: Mistral Cloud Chat Model.

Good to know

• Only files that change after the workflow is active are indexed. Files already in the folder are not read until you save them
  again.
• Prepare Embedding Document decodes the file as text from n8n's in-memory binary data, which is the default mode. PDFs and
  images come out as noise, and a filesystem or S3 binary mode gives the node a reference instead of the contents.
• The manual trigger goes to Set Variables with no file event, so a click on Test workflow does nothing useful.
• A changed file that has no stored points is skipped, not added. Set Variables also sets directory, which nothing reads.
• The line file created holds the time the workflow processed the file, not the time the file was made.
• The text of every file goes to Mistral for embedding, and each question sends the matching passages to Mistral. Qdrant keeps
  the passages. For bank statements or other private papers, check Mistral's data terms and where your Qdrant runs.
• Costs: embedding and answering both use your Mistral key.
• The chat has no filter. It searches every file in the collection.