Field notes from working through example 07 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

Ingestion is the unglamorous prep work for RAG. Done in four steps:

1. Parse the source (Markdown sections, code blocks, etc).
2. Chunk into ~500-1000 token pieces with overlap.
3. Generate embeddings for each chunk.
4. Insert into pgvector with metadata (source path, section, page).

What it looks like

func ingest(path string) error {
    raw, err := os.ReadFile(path)
    if err != nil { return err }

    sections := parseMarkdown(raw)
    var chunks []Chunk
    for _, s := range sections {
        chunks = append(chunks, splitWithOverlap(s, 800, 100)...)
    }

    embs, _ := embed.BatchGenerate(chunkTexts(chunks))

    for i, c := range chunks {
        _, _ = db.Exec(ctx,
            `INSERT INTO chunks (source, section, content, embedding)
             VALUES ($1, $2, $3, $4)`,
            path, c.Section, c.Text, embs[i])
    }
    return nil
}

What I learned

Chunk overlap is non-negotiable for prose. A 1000-token chunk that ends mid-thought loses the second half of the thought. 10-15% overlap means the next chunk picks up where the previous left off.

Metadata is what makes citations meaningful. Just storing chunks gets you “the answer is X.” Storing chunks + section + page lets you cite “the answer is X (from section 3.4, page 47).” The latter is what users trust.

Production connection

Genie’s document ingestion pipeline is structurally identical. The chunking heuristic is the same — token-aware, with overlap, section-anchored. The Ardan example is the cleanest implementation in idiomatic Go I’ve found.

Credit & reference. This post is field notes on example 07 from Ardan Labs’ Ultimate AI by Bill Kennedy + Florin Pățan, licensed Apache 2.0. The original example: cmd/examples/example07-ingestion/. 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.