Walkthrough · No. 02

GoPie: From a question to a checked report

How an AI data workspace turns tables and documents into answers, dashboards and reports, with the evidence attached. Shown on real CSR and BRSR filings.

CSR Insights chat on CSR spending in Northeast Indian states, showing the agent step trace, the generated SQL and the answer list.
Fig. 1 GoPie’s agent trace and SQL, as seen in CSR Insights. Factly product

§ 01 Problem

The problem

Numbers sit in spreadsheets, meaning sits in documents, and answers wait on whoever can write a query. Dashboards take a sprint and go stale; the monthly report is assembled by hand.

§ 02 Information and data

What made it hard

  1. Tables with stray spaces, cryptic headers and a dozen spellings of the same place.

  2. Codes without codebooks: state and district codes, scheme identifiers, abbreviations.

  3. Context in documents that never reach the analysis.

  4. Tables too large for a model to read; the system must query.

§ 03 Workflow and design

The decisions that shaped it

  • D1

    Clean on the way in

    Values and headers standardised on upload, code dictionaries applied (“US → United States”); the original untouched.

  • D2

    Query, do not read

    The agent reads the schemas, drafts SQL, runs and repairs it, and stops to explain if the data cannot answer.

  • D3

    Evidence attached to every answer

    Prose grounded in executed results; SQL, rows and document citations visible.

  • D4

    Dashboards in one click, reports you approve

    Dashboards generate from a dataset and are edited by chat; reports are approved before sharing.

  • D5

    Open core, your choice of hosting

    AGPL-3.0 core, self-hostable (Docker, Helm); the model provider is a configuration choice.

§ 04 Shipped, not planned

What it actually does

  • Upload CSV, Excel, Parquet or JSON, or connect PostgreSQL, MySQL, Snowflake, MotherDuck or Amazon S3
  • PDF, DOCX and notes as context documents beside the data
  • Answers with SQL, rows and document citations visible
  • Dashboards from a prompt, edited by chat; reports approved before sharing
  • Versioned REST API and SQL playground per dataset; MCP server; Slack, Teams and Google Chat bots

§ 05 Input → steps → output

Walkthrough

  1. Input

    Standardised datasets (a decade of Indian CSR filings from Dataful) and, for BRSR, the filing documents themselves.

  2. Step 1: Ask a question

    “What was the CSR expenditure in each state of Northeast India in 2024-25?” The agent plans and runs the query, step by step.

  3. Step 2: Check the working

    The generated SQL sits above the answer: the filter and grouping are there to check.

    CSR Insights chat on CSR spending in Northeast Indian states, showing the agent step trace, the generated SQL and the answer list.
    Fig. 2 Steps, SQL and answer in one view.
  4. Step 3: Bring in the documents

    In BRSR Insights the same engine answers from the data, then the filings, citing the PDFs.

    BRSR Insights chat answer with superscript citations and a "Sources cited" list of the company BRSR filing PDFs it drew on.
    Fig. 3 Structured data and documents in one cited answer (BRSR Insights).
  5. Step 4: Build a dashboard from a prompt

    “CSR spending across states over the years” becomes KPI tiles, a top-states list and a sector chart.

    CSR Insights AI-generated dashboard charts: top states by total CSR expenditure and CSR spending by sector.
    Fig. 4 A dashboard generated from one prompt.
  6. Step 5: Freeze a report

    A state-wise report for 2024-25: executive summary, charts, tables; reviewed, then exported to PDF.

    CSR Insights report "State-wise CSR Spending Report for 2024-25": the snapshot header with its generation date, Complete and Edited badges, and an Export PDF button.
    Fig. 5 A frozen, exportable narrative report.
  7. Output

    A dashboard and a report from questions, every number traceable to a query, reviewed before it leaves the workspace.

§ 06 Evaluation

How we check it

  • The agent’s limits are documented: tool-call budget per turn, sampled results for large tables, refusal when the data cannot answer.

  • Security and privacy behaviour are published: what the model sees, how credentials are encrypted.

  • In production as the engine behind CSR Insights, BRSR Insights and StateStats.

§ 07 Evidence

The figures

  • AGPL-3.0[1]

    open-source core

  • Go · LangGraph · Postgres · Qdrant · SvelteKit[2]

    published stack

  • 3[3]

    chat integrations: Slack, Microsoft Teams, Google Chat

§ 08 Limitations and lessons

Limitations and lessons

read this part first

Limitations

  • Data-privacy behaviour depends on the model gateway you choose; we configure it with you.
  • Generated dashboards can mislabel units or treat an aggregate as a category; review is part of the workflow.
  • Some questions need the data reshaped first.

Lessons from operation

  1. Showing the SQL changed the conversation: non-technical users trust it, technical users check it.
  2. Documents belong next to the data. Most “wrong” answers are right answers to a mis-defined column.
  3. Cleaning at ingest and good dictionaries beat model changes.

Sources (12)

  1. [1] factly/gopie (AGPL-3.0) · github.com (opens external site)
  2. [2] GoPie: home · gopie.ai (opens external site)
  3. [3] GoPie documentation: integrations · gopie.ai (opens external site)
  4. [4] GoPie for public data · gopie.ai (opens external site)
  5. [5] GoPie documentation: how the agent works · gopie.ai (opens external site)
  6. [6] GoPie documentation: data cleaning · gopie.ai (opens external site)
  7. [7] GoPie documentation: reports · gopie.ai (opens external site)
  8. [8] GoPie documentation: MCP server · gopie.ai (opens external site)
  9. [9] GoPie documentation: AI data privacy · gopie.ai (opens external site)
  10. [10] GoPie documentation: security overview · gopie.ai (opens external site)
  11. [11] CSR Insights: home · csr.dataful.in (opens external site)
  12. [12] BRSR Insights: home · brsr.dataful.in (opens external site)

Sources reviewed 15 November 2026.

Discuss a project

Have tables and documents that should answer questions?

Tell us about the data, the documents that explain it and the report you rebuild every cycle.