Field notes from working through example 30 of Ardan Labs’ Ultimate AI course by Bill Kennedy and Florin Pățan (Apache 2.0). My fork: PratikDhanave/ai-training. Thank you Bill and Florin for teaching this material — the patterns in this post are derived from the course; the production reflections at the end are mine.

What the example teaches

Docling extracts structured content from PDFs — text, tables, figures, headings — preserving layout. The example uses it as the parser stage of a RAG pipeline, then feeds the structured output through an LLM for summarisation / QA.

What it looks like

// Docling runs as a service (Python). The Go client hits its HTTP API.
client := docling.New("http://localhost:8081")
doc, _ := client.Extract(ctx, "annual-report.pdf")

for _, table := range doc.Tables {
    summary, _ := llm.Summarise(ctx, table.Markdown)
    storeChunk(doc.ID, table.PageNumber, summary)
}

What I learned

Tables are where most PDF pipelines fail. A bank statement, an invoice, a regulatory filing — the information density is in the tables. Naive text extraction loses the rows-and-columns relationship; LLM-based extraction is slow and unreliable. Docling preserves the structure.

Hybrid (parser + LLM) beats pure LLM for documents. Sending the raw PDF to a vision-capable LLM technically works. Costs 10× more in tokens. Docling’s symbolic extraction + LLM summarisation is the right split.

Production connection

The Genie XLSX + scanned-PDF loaders (pkg/loader/) follow the same pattern — symbolic extraction first, LLM as the narration layer. Tata client work also hit this — a quarterly report pipeline that switched from “send the PDF to Gemini” to “Docling first, summarise the structured output” cut token cost by 12× while improving table accuracy.

Credit & reference. This post is field notes on example 30 from Ardan Labs’ Ultimate AI by Bill Kennedy + Florin Pățan, licensed Apache 2.0. The original example: cmd/examples/example30-pdf-docling/. My fork with notes: PratikDhanave/ai-training. Highly recommend the course for anyone building AI applications in Go — the material is rigorous and the Kronk + yzma + llama.cpp pipeline gives you hardware-accelerated local inference end-to-end. Thank you, Bill and Florin.