Built by Anes Lachemi · Powered by n8n · LLM: Llama 4 Scout 17B via Groq · Vector DB: Supabase pgvector · Search: Apify · Image Gen: FLUX.1-schnell

Nami-AI is not a generic chatbot. It is an intelligent AI marketing consultant and entrepreneur advisor built specifically for the Algerian market.

It understands local consumer psychology, Algerian platforms (Ouedkniss, Yassir, BaridiMob, Temtem), wilaya-based targeting, Ramadan seasonality, cash-on-delivery dominance, and informal economy dynamics that no international AI tool accounts for.

Nami speaks Darija, French, Arabic, and English — automatically matching the user's language — and grounds every answer in DZ-specific context. It can also generate images directly in the chat via FLUX.1-schnell.

┌──────────────────────────────────────────────┐
│              Static Web UI                   │
│      index.html + script.js + style.css      │
└───────────────────┬──────────────────────────┘
                    │ POST FormData
                    ▼
┌──────────────────────────────────────────────┐
│            n8n Chat Webhook                  │
│   → Get User Prefs (Supabase)                │
│   → Format Prefs                             │
│   → Has File?                                │
│        YES → Extract PDF → File Agent        │
│        NO  →              Main Agent         │
└──────────────────────────────────────────────┘
                    │
       ┌────────────┴────────────┐
       ▼                         ▼
  Main Agent               File Agent
  • Groq Llama 4           • Groq Llama 4
  • RAG (Supabase)         • RAG (Supabase)
  • Apify Web Search       • Apify Web Search
  • FLUX Image Gen         • FLUX Image Gen

Google Drive Folder
      ↓
Loop Over Files → Filter (PDF / CSV / TXT)
      ↓
Download → Extract Text
      ↓
Embed with HuggingFace (768-dim)
      ↓
Insert into Supabase pgvector

Chat Trigger (public webhook, file upload enabled)
      ↓
Get User Prefs → Format User Prefs
      ↓
Has file attached?
  YES → Extract PDF → File Agent (Groq + RAG + Image Gen)
  NO  → Main Agent  (Groq + RAG + Web Search + Image Gen)

nami-ai/
├── README.md
├── index.html           ← Static chat UI
├── script.js            ← Frontend logic (fetch, state, rendering)
├── style.css            ← UI styles
├── workflow/
│   └── nami-ai.json     ← Exported n8n workflow (import into n8n)
└── supabase/
    └── setup.sql        ← SQL schema: documents table + match function

Run supabase/setup.sql in your Supabase project's SQL editor. It creates:

• documents table — stores embedded knowledge base chunks (768-dim vectors)
• user_preferences table — stores per-session user context
• match_documents() function — cosine similarity search via pgvector

In n8n → Settings → Credentials, add the following:

1. In n8n → Workflows → Import from file
2. Upload workflow/nami-ai.json
3. Assign your credentials to each node
4. In the "Search files and folders" node, replace the folder ID with your own Google Drive knowledge base folder ID

1. Upload .pdf, .csv, or .txt files to your Google Drive folder
2. In n8n, click "Test workflow" on the manual trigger (Pipeline 1)
3. The workflow loops through all files, embeds them via HuggingFace, and stores chunks in Supabase
4. Verify records appear in your documents table

⚠️ The ingestion pipeline does not deduplicate. Clear the documents table before re-running to avoid duplicate vectors.

1. Activate the workflow in n8n (toggle → ON)
2. Open the "When chat message received" node → copy the Production URL
3. In script.js, update PROXY_URL to your webhook production URL:

const PROXY_URL = 'https://your-n8n-instance.com/webhook/YOUR-WEBHOOK-ID/chat'

4. Deploy index.html, script.js, and style.css to any static host (GitHub Pages, Vercel, Netlify, etc.)

Both agents share the same core behavioral rules:

• Always query the knowledge base first before answering
• Respond in the user's exact language — Darija, French, Arabic, or English
• Never invent statistics, influencer names, or platform data
• Use web search only for time-sensitive queries (prices, news, trends)
• Simple questions → 1–3 sentences, no headers
• Marketing / business questions → numbered steps + one 💡 DZ Pro Tip
• Image requests → call FLUX tool once, return result immediately

Payment:   BaridiMob, CCP, Cash on Delivery (dominates e-commerce), Dahabia
Platforms: Ouedkniss, Yassir Business, Instagram/Facebook (primary), Telegram groups
Trust:     Phone number visibility, wilaya targeting, word-of-mouth (chka)
Seasons:   Ramadan (peak), Aïd el-Fitr, Aïd el-Adha, rentrée scolaire (Sept)
Barriers:  Low card penetration, delivery trust gaps, price sensitivity

Nami uses FLUX.1-schnell via the HuggingFace Inference Router. When a user requests an image in any language, the agent crafts a detailed English prompt and calls the tool.

Endpoint: https://router.huggingface.co/hf-inference/models/black-forest-labs/FLUX.1-schnell
Method:   POST
Headers:  Authorization: Bearer <YOUR_HF_TOKEN>
          Content-Type: application/json
          Accept: image/jpeg
Body:     { "inputs": "<detailed English prompt>" }

The interface is a fully static single-page application with no backend dependencies:

• Sidebar with chat history persisted in localStorage
• Welcome screen with suggested DZ-relevant prompts
• PDF file attachment support
• Markdown-lite rendering in bot responses
• Inline image rendering for generated images
• Mobile responsive with hamburger sidebar
• 90-second timeout for image generation with contextual loading message

To extend Nami:

1. Add new tools — Ouedkniss scraper, CRM integration, product price tracker
2. Expand the knowledge base — Algerian market research, sector reports, competitor analyses
3. Improve multilingual search — pre-process queries to append "Algérie" / "DZ" automatically
4. Add ingestion deduplication — hash-based check before inserting vectors
5. Automate session memory — write preferences back to Supabase at conversation end

This project is proprietary. All rights reserved by Anes Lachemi.