How to Build an AI Agent for Data Analysis, Visualization, AND Reporting (n8n)

Overview

This video walks through building a no-code AI agent system (using n8n) that turns large spreadsheets (thousands to tens of thousands of rows) into an interactive, natural-language data analysis + visualization + automated reporting tool—without writing code.

---

What the system does (end-to-end)

• Ingests spreadsheet data (CSV / Google Sheets / Excel-like) into a database (NoCodeDB / “noco db”).
• Connects that database to an n8n workflow so users can ask questions in plain English.
• The agent can:
  • Answer specific data questions, e.g., “What was the Department of Energy’s net cost in 2023?”
  • Generate detailed reports with summaries, interpretations, and conclusions
  • Create visual charts via QuickChart
    • Examples mentioned include bar charts
    • The approach supports multiple chart types (e.g., line / radar / bubble)
• Supports automation, such as generating scheduled reports (weekly/monthly) by instructing how prompts should be structured.

---

Demo outputs shown

• Query: “Department of Energy net cost in 2023”
  • Returns a breakdown including fields like gross cost, earned revenue, subtotal/change, net cost, and other financial metrics.
• Then: “Generate a visual”
  • Produces a bar chart and includes narrative insights (e.g., net cost significantly lower than gross cost).
• Then: “Turn this into an organized report”
  • Produces an executive-summary-style report with key figures, interpretations, and wrap-up conclusions.
• Additional example: “Department of State … covering the past 10 years”
  • Produces a multi-year trend report including:
    • An embedded chart
    • Sections such as financial overview, key figures (fiscal year, gross cost, earned revenue, net cost), trend analysis, and efficiency.

---

Tools / product features used

NoCodeDB (NoSQL-like database for spreadsheets)

• Import CSV into a database table
• Use API endpoints to query table data
• Requires:
  • API token
  • host (cloud recommended)
  • table ID
• The video emphasizes using API documentation to identify endpoints.

n8n

• Receives user chat input, calls NoCodeDB via an HTTP Request (GET method), and orchestrates further steps.
• Uses a “tools agent” pattern:
  • One workflow handles data querying + report generation
  • A separate tool/workflow handles chart generation

Claude (mentioned as the model used)

Used to:

• Help write prompt-related formatting
• Generate structured column lists
• Produce correct chart configurations in the required format
• Assist with filtering/query logic for correct agency matching

Note: GPT-4 might work similarly, but Claude was tested and preferred.

QuickChart

• A chart API that accepts chart configuration and returns an embeddable chart image/URL
• The workflow calls QuickChart with a chart.js-compatible configuration

---

Build guide / tutorial steps (key implementation points)

1. Prepare and import data

   • Start with a public spreadsheet/CSV (demo claims ~3,000 rows).
   • Claims support for 10k–20k+ rows.
   • Import into NoCodeDB, identify the table/base, and capture table ID.

2. Connect n8n to NoCodeDB using HTTP

   • Create an n8n HTTP Request (GET) call to NoCodeDB API endpoints.
   • Include:
     • workspace/base/table IDs
     • API token
     • host

3. Provide schema context to the agent

   • The agent needs the column names (e.g., record date, statement fiscal year, calendar year/quarter, agency name, earned revenue, net cost, etc.).
   • The video’s approach: use Claude to generate a stringified “columns” expression so the agent references schema accurately.

4. Agent “System Message”

   • Defines the agent as a data analysis expert for NoCodeDB table context.
   • For visualization tasks: requires returning a chart URL embedded using Markdown.

5. Memory handling

   • Uses window buffer memory in n8n.
   • Memory keyed by session ID from the chat input.

6. Tooling split: report workflow vs chart workflow

   • One workflow: query + narrative report generation
   • Another workflow (tool): chart configuration generation for QuickChart

7. Chart configuration generation

   • Chart workflow instructs the model to output valid chart.js configuration objects
   • Uses a structured output parser so the response matches chart generation needs
   • n8n then sends the config to QuickChart endpoint and returns the chart

8. Filtering / query correctness

   • The presenter notes initial issues where the agent selects the wrong agency records (e.g., mismatching “Department of Treasury”).
   • Fix involves iterating with Claude to craft a better filter/query expression for NoCodeDB searches (e.g., filter by agency name and relevant fields like earned revenue / net cost).

---

Review / analysis style claims

• The video frames the system as outperforming manual report writing and speeding up repetitive tasks.
• Emphasizes that AI + automations do the heavy lifting, reducing human effort for:
  • prompt input / scheduling instructions
  • supplying correct table IDs/schema
• Claims the system can be made “production ready” with minor tweaks.

---

Main speakers / sources

• Speaker: The video presenter (no name given in the subtitles)
• Referenced source: DK Chun (credited as providing inspiration via a prior video)
• Model/tool vendor mentioned: Claude (used for reasoning and structured outputs)