Build With Less AI

I Turned a 600-Page PDF Catalog Into a Product Database for Under $10

AI Export Lab · July 2026

**AI Export Lab** | July 2026...

**AI Export Lab** | July 2026

I am a TCM practitioner in China. Not a programmer. I ended up running a B2B auto parts export site — `jjradiator.com` — with fifteen thousand product pages, most of which started their life inside PDF catalogs. Some 300 pages. Some 600. One is 704.

For two years those PDFs sat on my hard drive, mocking me.

You cannot copy-paste them. You cannot import them into a spreadsheet. Data entry firms quoted me around $2,000 per catalog. OCR services quoted me by the page and returned garbage I still had to clean by hand. I actually paid one of them $180 for a 40-page sample, opened the output, and realized I would spend a full weekend fixing it. Then I would need to do that 15 more times.

I gave up on that plan.

What replaced it is the workflow in this post. One 600-page catalog into a structured database for **less than $10 in API costs**, plus one afternoon of me actually checking the output.

I am writing this so someone else with a hard drive full of PDFs and no data-entry budget can copy the approach.

---

The Stack

Four tools, working together:

No enterprise platform. No SaaS subscription in the hundreds. No dev team.

*Note for overseas readers: Doubao and Volcengine Ark need a mainland China phone number. Every deep-dive post below tells you what the overseas equivalent looks like — the same workflow runs on Qwen, Kimi, GLM, Xiaomi's overseas AI platform, DeepSeek's international API, or Western models. The pattern is portable. The specific brand names are not.*

---

The Five Things That Made It Work

Each of these is a separate deep-dive post. Skip to whichever one you need.

1. Pair a smart AI with a cheap one

Claude Code is good at understanding intent. DeepSeek V4 is good at chewing through volume for very little money. If you send everything to Claude, the bill creeps up fast. If you send everything to DeepSeek, you spend more of your own time back-and-forthing on script design.

The split I use: **Claude designs the script, DeepSeek runs inside it**.

→ Deep dive: *[Claude Code + DeepSeek V4: The Manager-Worker Pattern](/ai-export-lab/claude-code-deepseek-manager-worker-pattern/)*

2. Give your text AI eyes

DeepSeek cannot see images. Half of a product catalog lives inside images — OEM stamps, size tables rendered as pictures, part codes printed on photos.

The fix is small. Tell DeepSeek: whenever you need to read an image, call this vision API instead. One API key, one endpoint, one line of instruction. In China I use Doubao Vision Pro. Overseas equivalents work the same way.

→ Deep dive: *[How to Give DeepSeek Eyes: Adding Vision to a Text-Only Model](/ai-export-lab/how-to-give-deepseek-eyes-vision-api/)*

3. Cheap image recognition breaks the old cost math

I processed **6,000 product images for about $5**. Roughly $0.0008 per image. The last quote I got from a data entry firm for the same job was $1,400.

If you have images sitting in a folder that you've never processed because "it'd cost too much" — check the price today. It changed.

→ Deep dive: *[I Processed 6,000 Product Images for $5](/ai-export-lab/6000-images-for-5-dollars-doubao-vision-pro/)*

4. A monthly plan beats pay-as-you-go if you're using AI daily

I was spending about ¥15 (~$2) a day on DeepSeek V4 tokens. Switched to Volcengine Ark's ¥49/month plan. One full workday now consumes about 10% of the monthly allowance. Same work, monthly bill dropped to roughly a third.

The rule holds on other platforms too — Cursor, Claude Pro, ChatGPT Plus, OpenRouter. If your daily AI spend is above the price of a coffee, look for the flat plan.

→ Deep dive: *[The ¥49 AI Plan That Cut My DeepSeek Bill by Two-Thirds](/ai-export-lab/volcengine-ark-agent-plan-review/)*

5. Cut the big PDF into small PDFs before the AI touches it

This is the one that saved my project. My first attempt was: feed the whole 600-page PDF into one script, let it run overnight. It crashed at page 213. I fixed the error. It crashed at page 411 in a different way. The third attempt produced output that looked right at a glance but had silently swapped two brand columns for 80 pages.

I lost a week to that.

The rule I now use for every PDF job:

**Split the big PDF into small PDFs before the AI ever runs.** By brand. By category. By page range. Whatever natural division the PDF already has. A 600-page catalog becomes twenty 30-page PDFs. If one slice fails, I re-run one slice, not six hundred pages.

Three levels of splitting:

The AI is not smart enough to notice it's going off the rails at page 400. It is very good at handling page 12 of a 30-page slice.

**Big problem → many small problems → one careful check per small problem.**

---

The One Test That Decides Everything: Vector vs Image PDF

Before you commit to any workflow, do this:

**Open the PDF. Try to select some text with your mouse.**

If the text highlights and you can copy it → **vector PDF**. A script pulls the text out cleanly. Cost is negligible. My 600-page project was mostly this type.

If nothing highlights, or the page acts like a flat picture → **image PDF**. Every page has to go through OCR or vision AI. Cost goes up. Error rate goes up. Human review becomes non-optional.

| Type | Difficulty | Common source |

|------|------------|---------------|

| Vector PDF | Easy | Modern catalogs exported from CAD, InDesign, or Word |

| Text-based PDF | Medium | Digitized documents, older reports |

| Image PDF | Hard | Scanned paper catalogs, hand-drawn diagrams, photocopies |

Do this test first. I have wasted whole afternoons because I assumed a PDF was vector when it was actually a giant image.

---

The Boring Truth: Human Review Doesn't Go Away

The end of the pipeline is **not** "AI produces database, database goes on website."

The real end:

  1. AI produces a draft database.
  2. You spot-check a sample from every category — 10 rows each.
  3. Where you find mistakes, you feed them back to the AI and ask for a corrected pass.
  4. Repeat until the sample looks clean.
  5. Ship.

I usually catch two or three systematic errors per catalog. Once caught, one prompt tweak fixes thousands of rows. Miss it, and a customer emails you six weeks later saying "the OEM on your Volvo FH12 page is wrong" — and now you're debugging fifteen thousand rows.

Cheap AI plus disciplined review beats expensive AI plus blind trust. Every catalog I've done has proven this.

---

When This Approach Won't Work For You

I want to be honest about the failure cases, because I've hit all of them:

---

The Whole Thing In One Line

With a ¥49 monthly plan, ~$5 in vision API calls, Claude Code as the manager, DeepSeek V4 as the worker, and the discipline to split every big PDF into small ones — a non-programmer turned a 600-page paper catalog into a real product database.

Total out-of-pocket cost: under $10. Total elapsed time: one weekend.

If you have a stack of catalogs, or supplier docs, or scanned invoices, or old tender documents sitting on a hard drive — this is the year that story changes.

---

**What would you use this for?** If you have a specific PDF sitting on your drive you've been putting off, reply on the [contact page](/contact/) with a rough description. I'm curious what other people find in their own archives.

*This is a cornerstone post. Each of the five ideas above has its own deep dive linked inline. New posts every Monday and Friday at [buildwithlessai.com](https://buildwithlessai.com).*