Turn a folder of finance files into a fast Q&A helper. New, changed, or deleted files are kept in sync with a vector database so answers stay current. Teams can ask questions through a chat endpoint and get clear answers based on the latest documents.

The flow starts when the local folder changes. It reads the file, adds simple metadata like filename and month, splits the text into chunks, and creates embeddings with Mistral. Points are stored in Qdrant. If a file is updated or removed, matching points are found and deleted through the Qdrant API, then rebuilt when needed. A retriever pulls the best chunks for each question, and a Mistral chat model crafts the final reply. The chat trigger exposes a webhook so users can query the data.

Setup needs a folder path n8n can read, a Qdrant collection, and a Mistral API key. Expect faster research, fewer manual lookups, and better accuracy for finance reports and audits. Use it for month end close, expense reviews, compliance checks, and vendor inquiries. Follow the steps below to connect credentials, set the folder, and test the chat endpoint.

• Watches a local folder for add, change, and delete events
• Reads updated files and builds a clean text document with basic metadata like filename and month
• Splits large files into small chunks for better search and recall
• Creates embeddings with Mistral and writes vectors into a Qdrant collection
• Finds matching points in Qdrant by filename and removes them when files change or are deleted
• Retrieves the most relevant chunks from Qdrant for each user question
• Uses a Mistral chat model to turn retrieved text into clear answers
• Exposes a chat webhook endpoint for simple question submission
• Central config node to set directory path and collection name

• Reduce document lookup from hours to minutes by searching with natural language
• Automate syncing so files and vector data match 100 of the time
• Improve answer accuracy by keeping deleted or changed files updated in Qdrant
• Scale to thousands of pages with chunking and retrieval instead of manual reading
• Connect local files, Qdrant, and Mistral without custom code

1. Import the template into n8n: Create a new workflow in n8n > Click the three dots menu > Select 'Import from File' > Choose the downloaded JSON file.
2. You'll need accounts with Mistral and Qdrant. See the Tools Required section above for links to create accounts with these services.
3. In the n8n credentials manager, create a Mistral credential. Choose API Key, paste your key from the Mistral dashboard, and give the credential a clear name.
4. In the n8n credentials manager, create a Qdrant credential. Enter your Qdrant URL, port, and API key if enabled. Name the credential so you can find it later.
5. Open the Local File Trigger node and set the folder path that n8n can access. If running in Docker, mount the host folder into the n8n container and use that path. Keep awaitWriteFinish on for large files.
6. Open the Set Variables node and update the directory path and the Qdrant collection name you want to use.
7. Open the Qdrant related nodes and confirm they use your Qdrant credential. Make sure the collection name matches the value from Set Variables.
8. Check the HTTP Request nodes that search and delete points. Confirm the base URL points to your Qdrant server and the credential is selected.
9. Open the Mistral Embeddings and Mistral Chat Model nodes. Select your Mistral credential and choose the model you prefer.
10. Click Execute on the workflow and add a sample file to the watched folder. Confirm a new point appears in your Qdrant collection.
11. Update the same file and check that the matching point is removed and recreated. Delete the file and confirm the point is deleted.
12. Copy the Chat Trigger webhook URL from the node. Send a test question with curl or your browser and check the reply uses your document content.
13. If you see permission errors, verify the container volume mount and file paths. If you get 401 or 403 from Qdrant, recheck the API key and URL. If answers look off, adjust chunk size or add more metadata fields.