From Raw Data to Reports: Automating Business Intelligence with AI

In this guide, you will build an automated business intelligence pipeline that transforms raw data into formatted reports using NeuroLink. You will implement data ingestion, AI-powered analysis, chart generation, and report assembly – turning CSV files and database queries into executive-ready summaries with a single API call.

NeuroLink provides the building blocks for an end-to-end AI-powered BI pipeline: file processors for ingesting JSON, Excel, and other formats; a RAG CSV loader and chunking for datasets too large for a single context window; LLM generation with domain-specific prompts for analysis; per-call evaluation for report quality checks; and streaming output for real-time dashboards.

This tutorial walks through building a complete BI pipeline from scratch. By the end, you will have a system that takes raw data files, runs them through AI analysis, and produces quality-evaluated reports – on a schedule, without human intervention.

Architecture: AI-Powered BI Pipeline

The pipeline moves data through five stages: ingest, chunk, analyze, evaluate, and output.

flowchart LR
    DATA(["Raw Data<br/>CSV, JSON, Excel"]) --> PROC["File Processor<br/>Registry"]
    PROC --> CHUNK["RAG Chunking<br/>for large datasets"]
    CHUNK --> LLM["LLM Analysis<br/>with domain prompts"]
    LLM --> EVAL["Evaluation<br/>Accuracy + completeness"]
    EVAL --> REPORT(["Formatted Report<br/>Markdown / JSON"])

    style DATA fill:#3b82f6,stroke:#2563eb,color:#fff
    style PROC fill:#f59e0b,stroke:#d97706,color:#fff
    style LLM fill:#6366f1,stroke:#4f46e5,color:#fff
    style REPORT fill:#22c55e,stroke:#16a34a,color:#fff

Each stage is independent and testable. You can swap the file processor for a different data source, change the LLM provider, or adjust evaluation thresholds without touching the rest of the pipeline.

Step 1: Ingesting Data with File Processors

NeuroLink’s file-processing stack supports more than 260 file extensions. For BI pipelines, ProcessorRegistry can auto-select processors for Excel, JSON, XML, and other registered formats from a file’s MIME type and name; CSV and TSV ingestion is available through the RAG CSVLoader.

import { getProcessorRegistry } from '@juspay/neurolink/processors';

const registry = await getProcessorRegistry();

// Process Excel financial report
const financialData = await registry.processFile({
  id: 'financials',
  name: 'annual-financials.xlsx',
  mimetype: 'application/vnd.openxmlformats-officedocument.spreadsheetml.sheet',
  size: excelBuffer.length,
  buffer: excelBuffer,
});

// Process JSON API response
const apiData = await registry.processFile({
  id: 'metrics',
  name: 'metrics.json',
  mimetype: 'application/json',
  size: jsonBuffer.length,
  buffer: jsonBuffer,
});

if (!financialData?.success || !financialData.data) {
  throw new Error(financialData?.error?.userMessage ?? 'Excel processing failed');
}

if (!apiData?.success || !apiData.data) {
  throw new Error(apiData?.error?.userMessage ?? 'JSON processing failed');
}

The registry returns a common { success, data, error } envelope while each successful processor keeps a format-specific payload. Excel data exposes worksheets and rows, JSON exposes parsed and formatted content, and XML exposes parsed text and metadata. Handle CSV separately through NeuroLink’s RAG CSVLoader, as shown in the chunking stage below.

Note: For large Excel files with multiple sheets, the processor handles each sheet independently. You can target specific sheets by name or index if you only need a subset of the data.

Supported Data Formats

Format	MIME Type	Best For
CSV	text/csv	Sales data, logs, exports
Excel (.xlsx)	application/vnd.openxmlformats-officedocument.spreadsheetml.sheet	Financial reports, multi-sheet data
JSON	application/json	API responses, configuration data
XML	application/xml	Legacy system exports, feeds
TSV	text/tab-separated-values	Database exports, research data

Step 2: Chunking Large Datasets with RAG

When a dataset exceeds the LLM’s context window, you need to split it into manageable chunks. NeuroLink’s RAG chunking system handles this with configurable strategies.

// For datasets too large for a single LLM context window
// neurolink rag chunk sales-data.csv --strategy recursive --maxSize 2000

import { ChunkerRegistry, CSVLoader } from '@juspay/neurolink';

const salesDocument = await new CSVLoader().load('sales-data.csv', {
  outputFormat: 'text',
});
const chunker = ChunkerRegistry.get('recursive');
const chunks = await chunker.chunk(salesDocument.getContent(), {
  maxSize: 2000,
  overlap: 200,
  metadata: { source: 'sales-data.csv', type: 'sales-data' },
});
const salesDataContent = chunks.map((chunk) => chunk.text).join('\n\n');

console.log(`Split into ${chunks.length} chunks for analysis`);

The overlap parameter ensures that data points near chunk boundaries are not lost. For tabular data, set the overlap to include at least one complete row to avoid splitting a data record across chunks.

Chunking Strategy Selection

Different data types benefit from different chunking strategies:

• Recursive: Best for structured text with natural breakpoints (headers, sections). Use for report documents.
• Character: Simple fixed-size chunks. Use for unstructured text.
• Sentence: Splits on sentence boundaries. Use for narrative content.
• Token: Splits based on token count. Use when you need precise token budget control.

For tabular data like CSV, the recursive strategy with row-aware splitting works best. Set maxSize to fit comfortably within your model’s context window, leaving room for the system prompt and analysis instructions.

Step 3: LLM Analysis with Domain Prompts

The analysis step is where AI adds the most value. A well-crafted domain prompt transforms raw numbers into actionable insights.

import { NeuroLink } from '@juspay/neurolink';

const neurolink = new NeuroLink();

const analysisResult = await neurolink.generate({
  input: { text: `Analyze this sales data and identify key trends, anomalies, and actionable insights:\n\n${salesDataContent}` },
  provider: 'anthropic',
  model: 'claude-sonnet-4-5-20250929',
  systemPrompt: `You are a senior business analyst. When analyzing data:
1. Identify the top 3 trends with supporting data points
2. Flag any anomalies or outliers
3. Provide actionable recommendations
4. Use specific numbers and percentages
5. Format output as a structured report with sections`,
});

The system prompt is critical for analysis quality. Here are domain-specific prompts for common BI scenarios:

Sales Analysis Prompt

const salesAnalysisPrompt = `You are a senior sales analyst. Analyze the provided data and produce:

## Revenue Analysis
- Total revenue and month-over-month growth
- Top performing products/categories
- Revenue concentration risk (% from top 10 customers)

## Trend Identification
- Seasonal patterns
- Growth/decline trajectories
- Leading indicators

## Anomaly Detection
- Unusual spikes or drops (>2 standard deviations)
- Missing data periods
- Data quality issues

## Recommendations
- 3 specific, actionable recommendations
- Each with expected impact and implementation priority

Use specific numbers. Never say "significant increase" -- say "23% increase from $1.2M to $1.5M".`;

Financial Analysis Prompt

const financialAnalysisPrompt = `You are a CFO-level financial analyst. Analyze the provided data and produce:

## Financial Health
- Key ratios: current ratio, debt-to-equity, profit margins
- Cash flow analysis
- Working capital position

## Variance Analysis
- Budget vs actual for each line item
- Material variances (>5%) with root cause analysis

## Risk Assessment
- Concentration risks
- Trend-based projections
- Downside scenarios

Format as a board-ready executive summary. Lead with the conclusion, then support with data.`;

Multi-Chunk Analysis

For datasets split across multiple chunks, analyze each chunk independently and then synthesize:

async function analyzeMultiChunk(chunks: string[]): Promise<string> {
  // Step 1: Analyze each chunk independently
  const chunkAnalyses = await Promise.all(
    chunks.map((chunk, i) =>
      neurolink.generate({
        input: { text: `Analyze this data segment (${i + 1} of ${chunks.length}):\n\n${chunk}` },
        provider: 'anthropic',
        model: 'claude-sonnet-4-5-20250929',
        systemPrompt: 'You are a data analyst. Identify key metrics, trends, and anomalies in this data segment.',
      })
    )
  );

  // Step 2: Synthesize chunk analyses into a unified report
  const synthesis = await neurolink.generate({
    input: { text: `Synthesize these analyses into a single comprehensive report:\n\n${chunkAnalyses.map((a, i) => `--- Segment ${i + 1} ---\n${a.content}`).join('\n\n')}` },
    provider: 'anthropic',
    model: 'claude-sonnet-4-5-20250929',
    systemPrompt: `You are a senior business analyst. Combine segment analyses into a unified report.
Resolve any contradictions between segments. Identify cross-segment trends.
Produce a single, coherent executive summary.`,
  });

  return synthesis.content;
}

Step 4: Quality Evaluation

AI-generated reports must meet quality standards before distribution. NeuroLink’s per-call evaluation scores reports on relevance, accuracy, and completeness.

// Enable evaluation for this report-generation call
const evaluatedReport = await neurolink.generate({
  input: { text: `Analyze this sales data:\n\n${salesDataContent}` },
  provider: 'anthropic',
  model: 'claude-sonnet-4-5-20250929',
  enableEvaluation: true,
  evaluationDomain: 'business intelligence',
});

if (!evaluatedReport.evaluation) {
  throw new Error('Evaluation data was not returned');
}

const { relevance, accuracy, completeness, overall } =
  evaluatedReport.evaluation;

if (overall < 7) {
  throw new Error(`Report quality score ${overall} is below the threshold`);
}

Note: Choose quality thresholds from your own validation set and risk tolerance. Financial reports should require human review and independent verification; an evaluation score is not a substitute for checking the underlying calculations.

Domain-Specific Evaluation Criteria

With evaluationDomain set for analytics, review the returned scores using criteria such as:

• Accuracy: Are the numbers correct? Are calculations verifiable?
• Relevance: Does the analysis address the business questions?
• Completeness: Are all key metrics and dimensions covered?
• Insight quality: Are the insights actionable and specific?

For financial reports, you might add:

• Risk awareness: Does the report identify potential risks?
• Compliance: Does it follow regulatory reporting standards?
• Timeliness: Is the data current and the analysis timely?

Step 5: Streaming Output for Real-Time Dashboards

For dashboards that need to display AI insights as they are generated, use NeuroLink’s streaming API:

const result = await neurolink.stream({
  input: { text: `Generate a real-time executive summary from: ${latestMetrics}` },
  provider: 'openai',
  model: 'gpt-5.4',
});

for await (const chunk of result.stream) {
  if ('content' in chunk) {
    // Push to dashboard websocket
    dashboardSocket.send(chunk.content);
  }
}

Streaming is particularly valuable for long-form reports. Instead of waiting 30-60 seconds for a complete analysis, the dashboard can start rendering the executive summary within 1-2 seconds and progressively reveal details as they are generated.

WebSocket Integration

import { WebSocketServer } from 'ws';

const wss = new WebSocketServer({ port: 8080 });

wss.on('connection', (ws) => {
  ws.on('message', async (message) => {
    const { query, dataSource } = JSON.parse(message.toString());

    const data = await loadDataSource(dataSource);
    const result = await neurolink.stream({
      input: { text: `${query}\n\nData:\n${data}` },
      provider: 'openai',
      model: 'gpt-5.4',
      systemPrompt: salesAnalysisPrompt,
    });

    for await (const chunk of result.stream) {
      if ('content' in chunk) {
        ws.send(JSON.stringify({ type: 'analysis', content: chunk.content }));
      }
    }

    ws.send(JSON.stringify({ type: 'complete' }));
  });
});

Data Processing Sequence

Here is the full sequence for a scheduled daily report:

sequenceDiagram
    participant Sched as Scheduler
    participant NL as NeuroLink
    participant Proc as File Processor
    participant LLM as LLM Provider
    participant Eval as Evaluator
    participant Out as Report Store

    Sched->>NL: Trigger daily report
    NL->>Proc: Process CSV, Excel, JSON
    Proc-->>NL: Structured data
    NL->>LLM: Analyze with domain prompt
    LLM-->>NL: Analysis + insights
    NL->>Eval: Evaluate quality
    Eval-->>NL: Score >= 7 (pass)
    NL->>Out: Store formatted report

Multi-Source Report Aggregation

Production BI pipelines often combine data from multiple sources. Here is a pattern for ingesting and aggregating multiple data files into a single report:

async function generateAggregatedReport(
  dataSources: Array<{ id: string; name: string; buffer: Buffer; mimetype: string }>
): Promise<string> {
  const registry = await getProcessorRegistry();

  // Process all registered formats in parallel
  const processedData = await Promise.all(
    dataSources.map(source =>
      registry.processFile({
        id: source.id,
        name: source.name,
        mimetype: source.mimetype,
        size: source.buffer.length,
        buffer: source.buffer,
      })
    )
  );

  // Combine successful payloads with source labels
  const combinedContext = processedData
    .map((result, i) => {
      if (!result?.success || !result.data) {
        throw new Error(result?.error?.userMessage ?? `Could not process ${dataSources[i].name}`);
      }
      return `### Source: ${dataSources[i].name}\n${JSON.stringify(result.data)}`;
    })
    .join('\n\n---\n\n');

  // Generate unified analysis
  const report = await neurolink.generate({
    input: { text: `Analyze the following data from multiple sources and produce a unified report:\n\n${combinedContext}` },
    provider: 'anthropic',
    model: 'claude-sonnet-4-5-20250929',
    systemPrompt: `You are a senior business analyst working with multiple data sources.
Cross-reference data across sources for consistency.
Identify correlations between different datasets.
Produce a unified report with cross-source insights.`,
  });

  return report.content;
}

Template-Based Report Generation

For recurring reports with a consistent structure, use templates with placeholder injection:

const reportTemplate = `

## Executive Summary
{{executiveSummary}}

## Key Metrics
{{keyMetrics}}

## Trend Analysis
{{trendAnalysis}}

## Anomalies & Risks
{{anomalies}}

## Recommendations
{{recommendations}}

---
Generated by NeuroLink AI Pipeline | Data as of {{dataDate}}
`;


async function generateTemplatedReport(data: string, template: string): Promise<string> {
  // Generate each section independently for better quality
  const sections = ['executiveSummary', 'keyMetrics', 'trendAnalysis', 'anomalies', 'recommendations'];

  const sectionResults: Record<string, string> = {};

  for (const section of sections) {
    const result = await neurolink.generate({
      input: { text: `Generate the "${section}" section for a business report based on this data:\n\n${data}` },
      provider: 'anthropic',
      model: 'claude-sonnet-4-5-20250929',
      systemPrompt: `Generate only the "${section}" section. Be specific with numbers. Use bullet points for clarity.`,
    });
    sectionResults[section] = result.content;
  }

  // Fill template
  let report = template;
  
  report = report.replace('{{reportTitle}}', 'Q4 Business Review');
  report = report.replace('{{reportDate}}', new Date().toISOString().split('T')[0]);
  report = report.replace('{{dataDate}}', new Date().toISOString().split('T')[0]);

  for (const [key, value] of Object.entries(sectionResults)) {
    report = report.replace(`{{${key}}}`, value);
  }
  

  return report;
}

Production Pipeline Considerations

Scheduling

Use cron jobs or a queue system to trigger reports on a schedule:

# Daily report at 6 AM
0 6 * * * node /app/scripts/generate-daily-report.js

# Weekly summary every Monday at 8 AM
0 8 * * 1 node /app/scripts/generate-weekly-summary.js

# Monthly executive report on the 1st at 7 AM
0 7 1 * * node /app/scripts/generate-monthly-report.js

Versioning Reports

Store reports with timestamps and metadata for audit trails:

interface StoredReport {
  id: string;
  generatedAt: string;
  dataSourceVersions: Record<string, string>;
  model: string;
  evaluationScore: number;
  content: string;
}

Cost Management

For recurring reports, log input/output tokens, model, provider, and current unit pricing for each run. Aggregate those records by report type and schedule rather than relying on a static estimate, because provider prices and model choices change over time.

What’s Next

You have completed all the steps in this guide. To continue building on what you have learned:

1. Review the code examples and adapt them for your specific use case
2. Start with the simplest pattern first and add complexity as your requirements grow
3. Monitor performance metrics to validate that each change improves your system
4. Consult the NeuroLink documentation for advanced configuration options

---

Related posts:

• Structured Output: JSON Schema Enforcement with NeuroLink
• Batch Processing: Thousands of AI Responses Efficiently
• AI Streaming in React: Server-Sent Events and WebSockets