What problem does Reports That Write Themselves solve?

Monday-morning reports that write themselves: SQL pulls the numbers, RAG pulls the narrative context, an LLM composes commentary that is checked against the figures it cites, and the result lands in inboxes as a formatted document. The division of labor is the whole design: numbers come from the database, never from the model.

How does Reports That Write Themselves work at a high level?

A report automation that combines structured and unstructured sources under a strict division of labor. Scheduled SQL queries against warehouse views produce every number in the report; retrieval over meeting notes, tickets, and documents supplies the qualitative context; and an LLM composes the narrative — but may only reference figures by ID from the SQL results, never generate them. A verification pass checks tha…

When should a team implement Reports That Write Themselves?

The model never does arithmetic: every figure in the report is computed by SQL and referenced by ID, not regenerated as tokens.

What should teams remember about the model never does arithmetic?

The model never does arithmetic: every figure in the report is computed by SQL and referenced by ID, not regenerated as tokens.

What should teams remember about narrative context comes from retrieval over real sources?

Narrative context comes from retrieval over real sources — meeting notes, tickets, docs — so commentary explains the numbers instead of decorating them.

What should teams remember about a verification pass compares prose against bound figures before anything ships; mismatches block the report, not the reader?

A verification pass compares prose against bound figures before anything ships; mismatches block the report, not the reader.

What should teams remember about anomalies get first?

Anomalies get first-class treatment: significant deviations from trend trigger a deeper drill-down section automatically.

What should teams remember about the same pipeline serves weekly ops reports, monthly board packs, and client?

The same pipeline serves weekly ops reports, monthly board packs, and client-facing summaries — only the views and templates change.

What constraint or design note matters for Reports That Write Themselves?

Design rule with no exceptions: if a number appears in the report, a SQL query produced it and a figure ID binds it. A model that generates '$1.2M' as tokens is a typo generator with confidence.

How does Reports That Write Themselves work step by step?

Metrics come from versioned warehouse views: each report section maps to named queries with owners — the report's numbers are as governed as any dashboard's. Every query result becomes a figure record: value, unit, period, comparison to prior period, and a stable figure ID the narrative must reference.

What happens during "step 1" in the how it works stage of Reports That Write Themselves?

Metrics come from versioned warehouse views: each report section maps to named queries with owners — the report's numbers are as governed as any dashboard's.

What happens during "step 2" in the how it works stage of Reports That Write Themselves?

Every query result becomes a figure record: value, unit, period, comparison to prior period, and a stable figure ID the narrative must reference.

What happens during "Anomaly detection runs before writing" in the how it works stage of Reports That Write Themselves?

Anomaly detection runs before writing: deviations beyond trend bands get flagged, and flagged metrics automatically earn a drill-down query set — the report investigates itself before commenting.

What happens during "Context retrieval targets the flags" in the how it works stage of Reports That Write Themselves?

Context retrieval targets the flags: for each anomaly, retrieval searches meeting notes, tickets, and release logs from the period for candidate explanations — with citations.

What happens during "step 5" in the how it works stage of Reports That Write Themselves?

The LLM composes narrative with hard constraints: figures referenced only by ID ({{fig:mrr_delta}}), qualitative claims only with a retrieval citation, tone and structure from the section template.

What happens during "Verification is mechanical" in the how it works stage of Reports That Write Themselves?

Verification is mechanical: every figure token in the rendered prose is compared to its bound value; every citation is checked to exist; failures loop back to recompose, never ship as-is.

What happens during "Render and distribute" in the how it works stage of Reports That Write Themselves?

Render and distribute: templated document (PDF, email, or doc), consistent layout week over week — readers scan the same shape every time, which is half the value of a report.

How does figure binding prevent numeric hallucinations in reports?

The workflow uses a structured JSON record so later stages can validate routing, provenance, and audit decisions mechanically instead of trusting free-form model output.

What role does scheduling play in Reports That Write Themselves?

Scheduling — cron trigger with on-demand re-runs per section

What role does metrics play in Reports That Write Themselves?

Metrics — read-only warehouse views (Postgres / BigQuery-class), versioned in the repo

What role does context play in Reports That Write Themselves?

Context — retrieval index over meeting notes, tickets, release logs (see the document intake pipeline)

What role does composition play in Reports That Write Themselves?

Composition — LLM with templated sections and figure-ID placeholders; structured output

What role does verification & render play in Reports That Write Themselves?

Verification & render — deterministic binding checker, then HTML-to-PDF or email template

How does Reports That Write Themselves handle a source query fails?

A source query fails → the section renders with an explicit 'data unavailable' block; the report ships on time with a visible hole, never a silently stale number.

How does Reports That Write Themselves handle no context found for an anomaly?

No context found for an anomaly → the narrative says 'no documented explanation found in this period's notes' — an honest gap that often triggers the right conversation.

How does Reports That Write Themselves handle late-arriving data revises a figure?

Late-arriving data revises a figure → re-run regenerates the report as v2 with a change note; distributed versions are immutable.

How does Reports That Write Themselves handle conflicting context (two notes explain the dip differently)?

Conflicting context (two notes explain the dip differently) → both candidate explanations ship with citations, mirroring the research-brief conflict rule.

How does Reports That Write Themselves handle a reader replies asking a follow-up?

A reader replies asking a follow-up → the question routes into an NL2SQL lane against the same views, answering from the same governed numbers.

What is Figure binding?

The contract that every number in generated prose references a computed figure record by ID rather than being generated as text. A mechanical verification pass compares the rendered prose against bound values before distribution — making numeric hallucination structurally impossible rather than merely discouraged.

How does Reports That Write Themselves apply NL2SQL?

Reports That Write Themselves treats NL2SQL as a first-class design constraint rather than an afterthought in orchestration.

How does Reports That Write Themselves apply Report Generation?

Reports That Write Themselves treats Report Generation as a first-class design constraint rather than an afterthought in orchestration.

What is this automation workflow about?

A report automation that combines structured and unstructured sources under a strict division of labor. Scheduled SQL queries against warehouse views produce every number in the report; retrieval over meeting notes, tickets, and documents supplies the qualitative context; and an LLM composes the narrative — but may on…

Who should read Reports That Write Themselves: SQL for the Numbers, RAG for the Narrative?

Engineers, hiring managers, and technical leaders interested in Data Pipeline, NL2SQL, Reporting, RAG, LLM, ai-engineering, rag-systems, enterprise-automation.

What are the key takeaways from Reports That Write Themselves: SQL for the Numbers, RAG for the Narrative?

Readers will learn about The model never does arithmetic: every figure in the report is computed by SQL and referenced by ID, not regenerated as tokens., Narrative context comes from retrieval over real sources — meeting notes, tickets, docs — so commentary explains the numbers instead of decorating them., A verification pass compares prose against bound figures before anything ships; mismatches block the report, not the reader., Anomalies get first-class treatment: significant deviations from trend trigger a deeper drill-down section automatically., The same pipeline serves weekly ops reports, monthly board packs, and client-facing summaries — only the views and templates change..

How does Reports That Write Themselves: SQL for the Numbers, RAG for the Narrative relate to Dhruvil Patel's work?

Reports That Write Themselves: SQL for the Numbers, RAG for the Narrative reflects hands-on work in RAG systems, multi-agent workflows, and production AI infrastructure.

Automated Report Generation: SQL Numbers, RAG Context, Verified Narra…

Ask an LLM to write your weekly report from raw data and you'll get beautiful prose with numbers that are almost right — which is the worst kind of wrong, because almost-right numbers get forwarded to boards. The fix isn't a better model. It's refusing to let the model touch arithmetic at all.

The pipeline

How it works

Metrics come from versioned warehouse views: each report section maps to named queries with owners — the report's numbers are as governed as any dashboard's.
Every query result becomes a figure record: value, unit, period, comparison to prior period, and a stable figure ID the narrative must reference.
Anomaly detection runs before writing: deviations beyond trend bands get flagged, and flagged metrics automatically earn a drill-down query set — the report investigates itself before commenting.
Context retrieval targets the flags: for each anomaly, retrieval searches meeting notes, tickets, and release logs from the period for candidate explanations — with citations.
The LLM composes narrative with hard constraints: figures referenced only by ID ({{fig:mrr_delta}}), qualitative claims only with a retrieval citation, tone and structure from the section template.
Verification is mechanical: every figure token in the rendered prose is compared to its bound value; every citation is checked to exist; failures loop back to recompose, never ship as-is.
Render and distribute: templated document (PDF, email, or doc), consistent layout week over week — readers scan the same shape every time, which is half the value of a report.

Figure binding shape

{
  "figure_id": "fig:mrr_delta",
  "query": "weekly_mrr_delta_v3",
  "value": -4.2,
  "unit": "percent",
  "period": "2026-W27",
  "vs_prior": { "value": 1.8, "direction": "down_from_up" },
  "anomaly": true,
  "drilldown": ["churn_by_plan", "downgrades_by_cohort"],
  "context_citations": ["note_0712_pricing_meeting", "ticket_8841"]
}

Stack

Scheduling — cron trigger with on-demand re-runs per section
Metrics — read-only warehouse views (Postgres / BigQuery-class), versioned in the repo
Context — retrieval index over meeting notes, tickets, release logs (see the document intake pipeline)
Composition — LLM with templated sections and figure-ID placeholders; structured output
Verification & render — deterministic binding checker, then HTML-to-PDF or email template

A good automated report doesn't just say revenue dipped 4.2%. It says it dipped 4.2%, that the dip is concentrated in one plan tier, and that the pricing change discussed in Tuesday's meeting is the likely reason — with links to all three.

Reports That Write Themselves: SQL for the Numbers, RAG for the Narrative

Key Takeaways

FAQ

The pipeline

How it works

Figure binding shape

Stack

Key Takeaways

FAQ

Related Expertise

Related Concepts

Related Projects

Related Research

Related Articles

Related Automations

The pipeline

How it works

Figure binding shape

Stack