n8n RAG System for Rollout (Company Policy) Documentation

This project provides n8n workflows that watch a Google Drive folder for new or updated company policy documents, index them into a vector store, and answer employee questions via a webhook using retrieved context.

What's in this repo

Item	Purpose
n8n - Rollout/workflows/rollout-drive-watch.json	Runs on a schedule; lists files in your policy folder that are new or modified since the last run; triggers the ingest workflow for each file.
n8n - Rollout/workflows/rollout-ingest.json	Triggered by the watch workflow (or manually for backfill): downloads each file, extracts text, chunks, embeds, and upserts into the vector store.
n8n - Rollout/workflows/rollout-policy-ask-workflow.json	Webhook for employee questions: retrieves relevant chunks from the vector store and returns a grounded answer (with optional Structured Governor).
n8n - Rollout/docs/	Internal notes (chunking, folder ID, etc.).

Prerequisites

1. n8n instance (cloud or self-hosted).
2. Google Drive
   • A dedicated folder for Rollout policy docs (signage rules, feedouts, employee/manager expectations).
   • Google Drive OAuth2 credentials in n8n (same as "Google Drive" or "Google API").
3. OpenAI
   • OpenAI API key for embeddings (and for the Q&A LLM).
4. Vector store
   • Qdrant (Qdrant Cloud or self-hosted) and Qdrant credentials in n8n, or
   • Supabase with pgvector and corresponding credentials in n8n.

Setup

1. Credentials in n8n

• Google Drive: Create OAuth2 credentials (e.g. "Google Drive OAuth2") and complete the consent flow so n8n can list and download files from your policy folder.
• OpenAI: Create an API key credential in n8n.
• Qdrant (or Supabase): Create the credential and note the collection name (e.g. rollout_policy).

2. Create the vector store collection

• Qdrant: Create a collection (e.g. rollout_policy) with the same dimension as your embedding model (e.g. 1536 for text-embedding-3-small, 3072 for text-embedding-3-large). You can create it via Qdrant UI or let the first ingest run create it if the node supports auto-creation.
• Supabase: Create a table with a vector column and use the same collection/table name in the workflows.

3. Import the workflows

1. In n8n: Workflows → Import from File (or paste JSON).
2. Import in this order (so IDs and names match):
   • n8n - Rollout/workflows/rollout-ingest.json
   • n8n - Rollout/workflows/rollout-drive-watch.json
   • n8n - Rollout/workflows/rollout-policy-ask-workflow.json
3. In rollout-drive-watch:
   • Set the Google Drive node to your credential and set the folder ID (see Finding your folder ID).
   • In the Execute Workflow node, select the rollout-ingest workflow so each new/updated file triggers ingestion.
4. In rollout-ingest:
   • Attach your Google Drive, OpenAI, and Qdrant (or Supabase) credentials to the relevant nodes.
   • Set the vector store collection name to match your collection (e.g. rollout_policy).
5. In rollout-policy-ask-workflow:
   • Attach OpenAI and Qdrant (or Supabase) credentials.
   • Set the same collection name as in the ingest workflow.

4. Activate

• Activate rollout-drive-watch so it runs on the schedule (e.g. every 30–60 minutes).
• Activate rollout-policy-ask-workflow so the webhook is live.

Finding your folder ID

Open your policy folder in Google Drive in the browser. The URL looks like:

https://drive.google.com/drive/folders/FOLDER_ID

Use FOLDER_ID in the rollout-drive-watch Google Drive node (e.g. in the "Folder" or "Search" query).

Webhook: asking questions

Endpoint: POST /webhook/ask-rollout-policy (or the path shown in your n8n Webhook node).

Request body (JSON):

{
  "message": "What are the current signage rules for the break room?"
}

Alternatively you can send user_message or text instead of message.

Response (200):

{
  "success": true,
  "answer": "...",
  "sources": [
    { "file_name": "Signage Policy 2024.pdf", "snippet": "..." }
  ]
}

On error you get success: false and an error field; status code 400 for bad request, 502 if the LLM or vector store fails.

Optional: Structured Governor

To keep answers factual and low-drift (like the diet feedout flow), you can run the Structured Governor on the same server and add an HTTP Request node in rollout-policy-ask-workflow that calls POST http://localhost:5050/govern with the user message and RAG context, then return the governor's resolution.answer. See the Structured Gov README for running the governor and the diet-feedout-ask workflow for the payload shape.

File format support

Format	Supported	Notes
Google Docs	Yes	Export as HTML/plain text; template uses Markdown conversion where applicable.
PDF (text-based)	Yes	Use Extract From File on the downloaded binary.
DOCX	Yes	Use Extract From File (or a DOCX community node) on the downloaded binary.
Scanned PDFs	No (v1)	Would require OCR; add later if needed.

Reindex / backfill

To index all existing files in the folder once:

1. Run rollout-drive-watch once with the "Code" node temporarily changed so lastRun is set to a very old date (e.g. '1970-01-01T00:00:00Z'), so the Search returns all files.
2. Or create a one-off workflow that lists all files in the folder (no date filter) and executes rollout-ingest with that list.

Troubleshooting

• No files found: Check folder ID and that the Google Drive credential has access to that folder. Ensure lastRun in the Code node is not in the future.
• Ingest fails on a file: Check file type (Google Docs vs binary). For Google Docs, ensure the node is set to export (e.g. HTML/text). For PDF/DOCX, ensure Extract From File receives the binary and the correct MIME type.
• Empty or poor answers: Confirm the vector store collection has data (run ingest first). Increase "top k" or limit in the Q&A workflow retrieval step. Check that chunk size and overlap in the ingest pipeline suit your policy doc length.

License

Same as the rest of your n8n Rollout project.