<aside>

Goal:

Let non-technical users ask sales questions in plain English and get data-backed insights drawn from Supabase (PostgreSQL).

Problem:

Analysts lose time translating natural-language questions into SQL and formatting results. Stakeholders want quick, readable answers (not tables or code).

Solution:

An n8n workflow that converts the user’s question into SQL, executes it on the “Sales Data 2024/2025” tables, and returns a narrative insight with key numbers highlighted.

Step 1: Trigger (Chat UI)

• Node: When chat message received — starts the flow when a user submits a question.

Step 2: Generate SQL from Natural Language

• Node: SQL Generator (LLM agent) — converts the question to strict PostgreSQL using rules (quote identifiers, no inline ORDER BY in UNION parts, use the existing “Profit ($)” column, etc.).
• Model & Memory: Powered via OpenRouter with a short-term Simple Memory to keep chat context.

Step 3: Execute SQL

• Node: Execute a SQL query — runs the generated query against Postgres/Supabase.

Step 4: Normalize Output

• Node: Combine all data — wraps SQL rows into a single JSON payload (sql_output) for the responder agent.

Step 5: Create the Narrative Answer

• Node: Response Agent (LLM) — receives sql_output + the original user message and writes a concise, number-rich summary (2–3 short paragraphs) formatted for business users.
• Model: OpenRouter Chat Model dedicated to response generation.

Tools Used:

• n8n (LangChain chat trigger + agents, Postgres node)
• Supabase / PostgreSQL (sales datasets for 2024 & 2025) — queried via the Postgres node.
• OpenRouter LLMs (separate models for SQL generation and narrative response).

Impact:

• Stakeholders ask “What was the most expensive product in 2025?” and get a clear answer with exact figures — no SQL knowledge required.
• Consistent guardrails (e.g., always use Profit ($) from the table) reduce logic errors and rework.

🧠 Learnings:

• Splitting responsibilities (SQL agent → DB → response agent) increases reliability and makes each part easier to debug.
• Normalizing DB rows into a single JSON block simplifies prompt design for the responder agent.
• Lightweight chat memory helps follow-ups (“Now show me profits by region”) while keeping the SQL agent deterministic.

</aside>

📂 Download & Explore the Workflow:

Sales Insights Chatbot JSON.json

View GitHub Repository :

Sales-Insights-Chatbot-n8n/Sales Insights Chatbot JSON.json at main · AlvLeoAI/Sales-Insights-Chatbot-n8n