WhatsApp RAG Agent (Text + Voice) with Weekly Google Drive Sync

One-line summary : Answers WhatsApp in under 100 words, understands voice notes, and retrieves trusted answers from your Google Drive docs (RAG) kept fresh weekly.

What this template does

• Reply to WhatsApp messages in a polite, human tone with ≤100 words.
• Understands text and voice notes: auto-downloads audio and transcribes to text.
• Retrieves answers from your knowledge base (RAG): Google Drive docs → chunk → embed → store in Supabase → rerank with Cohere.
• Keeps short-term memory across the conversation to avoid repetition.
• Weekly doc sync from a selected Google Drive folder so non-technical staff can update content without touching n8n.

Why it matters

• Zero busywork: Update a Google Doc; the bot learns it on the next sync.
• Trustworthy answers: Responses come from your vetted docs, not random web text.
• Voice-first friendly: Handles the WhatsApp reality of “send a voice note.”
• Safer by design: Guardrails—no pricing unless in KB, no pushy sales, no medical advice.

Triggers

• WhatsApp Trigger: Receives incoming messages (text or audio).
• Google Drive Trigger (weekly): Detects new/updated files in the chosen folder for ingestion.

App credentials required

• WhatsApp Business Cloud (App ID, Token, Phone Number ID)
• OpenAI (Chat + Embeddings)
• Cohere (Rerank)
• Supabase (SUPABASE_URL, ANON_KEY)
• Google Drive (OAuth2) + target Folder ID

Suggested environment variables

WHATSAPP_APP_ID=
WHATSAPP_TOKEN=
WHATSAPP_PHONE_NUMBER_ID=
OPENAI_API_KEY=
COHERE_API_KEY=
SUPABASE_URL=
SUPABASE_ANON_KEY=
GDRIVE_FOLDER_ID=
RAG_TABLE_NAME="documents"     # table to store vectors/metadata
MAX_ANSWER_WORDS=100            # guardrail for concise replies

Architecture overview

• Answer-time lane (RAG Tool):

  1. Receive WhatsApp message → 2) Transcribe audio if present → 3) Maintain short-term memory → 4) Retrieve from Supabase vectors (topK) → 5) Rerank with Cohere → 6) Compose ≤100-word reply with sources → 7) Send on WhatsApp.

• Ingest lane (Weekly Sync):
  A) Detect Drive file updates → B) Download & convert (Docs → text/plain) → C) Chunk (size/overlap) → D) Embed with OpenAI → E) Upsert to Supabase with metadata & hashes.

How it works (node rundown)

Notes

• The KB Tool is the combo of steps 8–9–7 at answer time (retrieve → rerank → answer).
• The Ingest lane is steps 11–17 (weekly sync of your Drive folder).

Setup (7‑minute sprint)

1. Import the workflow JSON.

2. Connect credentials: WhatsApp, OpenAI, Cohere, Supabase, Google Drive.

3. Google Drive Trigger: paste your Folder ID; keep fileUpdated event.

4. Download File: ensure Google Docs convert to text/plain.

5. Supabase Vector Store (insert): set table name to documents (or your schema).

6. Character Text Splitter: keep chunkSize=2000, overlap=300 (balanced recall/latency).

7. Retrieve-as-tool: set topK=10 and enable reranker.

8. Send WhatsApp Message mapping:

   • Recipient: {{$("WhatsApp Trigger").item.json.messages[0].from}}
   • Body: {{$json.output}}

9. Test:

   • Send a text and a voice note to your WhatsApp number → confirm concise answers.
   • Drop a Google Doc into the watched folder → verify it’s chunked/embedded on the next weekly poll (or run ingest nodes once manually).

Prompt, tone & guardrails

• System prompt:

  • Be polite, human, and concise (≤ MAX_ANSWER_WORDS).
  • Cite or reference only the KB content; if unknown, say so and offer to escalate.
  • No prices unless present in the KB. No medical advice.

• Temperature: start at 0.2 for factual replies.

• Memory window: keep a short rolling buffer (e.g., last 4–6 turns).

Data model (minimum viable)

Table documents (example columns):

• id (uuid)
• source_url (text)
• title (text)
• chunk (text)
• embedding (vector/float[] depending on extension)
• chunk_hash (text)
• updated_at (timestamp)

Indexes

• Unique index on chunk_hash to dedupe
• Index on updated_at for syncs

Observability & ops

• Log question, selected chunk ids/hashes, and final response to a DB/Sheet for QA.
• Add a low-confidence route (score threshold) → Slack/Telegram escalation to a human.
• Track latency and token usage to tune topK and chunk sizes.

Customization

• Latency vs quality: try topK=6–8 and chunkSize=1200 for speed.
• Languages: ASR node can be swapped for native multilingual output.
• Escalation: add channel handoff on low confidence or no KB hits.
• Sync cadence: change Drive Trigger to daily if content updates frequently.

Safety & compliance

• No medical advice. If uncertain or clinical, ask to schedule a consult or refer to the right department.
• PII: Don’t log full phone numbers in plaintext analytics; hash where possible.
• Prices & rates: Only answer if present in the KB; otherwise hand off to front desk.

Troubleshooting

• No reply sent: Ensure Send message node reads {{$json.output}} (Agent’s response property).
• Audio path failing: Confirm attachments[0].data_url exists and HTTP node fetches a valid file.
• KB not updating: Manually execute the ingest lane; check rows in Supabase.
• Irrelevant answers: Lower temperature to 0.2, increase overlap to 400, and verify Drive docs are clean and structured.

Categories & tags

• Categories: AI, Customer Support, Healthcare Ops, RAG, WhatsApp
• Tags: WhatsApp, RAG, Google Drive, Supabase, OpenAI, Cohere, Voice Notes

Pricing (rough, BYO keys)

• n8n: self-host free; n8n.cloud billed by plan.
• OpenAI, Cohere: usage-based by tokens/calls.
• Supabase: free tier + usage; vector storage billed by size.
• WhatsApp Cloud: Meta pricing per conversation.

Nodes used in workflow

WhatsApp Trigger, Switch, HTTP Request, OpenAI (ASR + Chat + Embeddings), Merge, Simple Memory, Supabase Vector Tool (retrieve), Cohere Reranker, Google Drive Trigger, Set, Google Drive – Download, Character Text Splitter, Default Data Loader.