AI Agent Bot that understands Text, Audio, Image and Documents

Create your Custom Business AI Agent that speaks, sees, listens and replies to your customers.

🚀 What this workflow does

1. Receives any inbound WhatsApp message via a PulpoChat Trigger
2. Detects the medium – text, voice note, image or document (PDF)
3. Processes accordingly
   • Text → straight to the AI brain
   • Voice notes → download ➜ Whisper transcription
   • Images → download ➜ GPT-4o Vision analysis
   • PDFs only → download ➜ text extraction
4. Feeds the cleaned input + short-term memory buffer (20 turns) to an OpenAI Chat Agent (GPT-4o-mini by default)
5. Sends the answer back through PulpoChat:
   • If the user sent audio, the bot replies in audio (OpenAI TTS ➜ saves mp3 to Google Drive ➜ returns the public link).
   • Otherwise, returns plain text.
6. Gracefully rejects anything that isn’t text, image, audio or a PDF (“Sorry, you can only send …”)

Result: a polite, context-aware concierge that can read your contract, describe your cat photo, or summarize a 3-minute rant into a single line—without ever leaving WhatsApp.

🧩 Key components

Node	Purpose
PulpoChat Trigger / PulpoChat	Receive & send WhatsApp messages
Switch → “Input type”	Routes to Text / Audio / Image / Document branches
HTTP Request	Securely downloads media from PulpoChat
OpenAI Whisper	Turns voice notes into text
GPT-4o Vision	Describes images in detail
Extract From File	Converts PDFs to text
LangChain Agent	Central brain with custom system prompt
Memory Buffer Window	Keeps the last 20 turns per chat
OpenAI TTS (“Generate Audio Response”)	Converts answers to speech (voice “nova”)
Google Drive (Upload + Delete)	Stores the mp3, grabs a share link, cleans up

(Sticky notes in the canvas label the four media lanes so future-you won’t get lost.)

🛠️ Prerequisites

• PulpoChat device + API key
• OpenAI API key (chat, whisper, TTS, vision)
• Google Drive OAuth credentials (for audio replies)

💡 Ideas & extensions

• Pipe extracted conversation data into HubSpot or Airtable.
• Replace GPT-4o with your on-prem model ➜ just swap the Chat node.
• Add a Sentiment node to auto-escalate angry customers.
• Expand document branch to Word, PowerPoint or spreadsheets.

⚖️ Limits & best-practice nudges

• Only PDFs are accepted for now; other file types trigger a polite rejection.
• The workflow rate-limits itself by design (single execution per message), but you may want extra guards if you point it at a large audience.
• Delete Google Drive files after sending (already included) to keep storage costs clean.
• Remember WhatsApp’s 24-hour customer-initiated window.

🏁 Ready, set, automate!

Import → Hit Active. Your WhatsApp number just became a futuristic, multimodal AI agent. Enjoy the peace and quiet while it handles the chatter. 😉