Beginner's mistake I made: Not cleaning PDFs before upload. Here's what happened.

Hey folks, I just spent a frustrating afternoon debugging some truly bizarre and inconsistent answers from NotebookLM, and the root cause was entirely my fault. I wanted to share this so you can avoid the same CI/CD-for-your-brain pipeline issue 😅

I was uploading a bunch of technical PDFs (API docs, some academic papers) to create a source corpus. The answers I got back were sometimes brilliant, sometimes wildly off—like hallucinating tool names or misattributing concepts. I finally realized the problem: I hadn't "cleaned" my PDFs before upload. Many PDFs, especially those with multi-column layouts, complex headers/footers, or generated from slides, create absolute chaos when extracted as plain text.

Here's a comparison of a raw PDF text extraction vs. a cleaned version, using a simple `pdftotext` command:

```bash
# Raw extraction (messy, multi-column blended)
pdftotext my_doc.pdf output_raw.txt

# Cleaned with some flags (often better layout preservation)
pdftotext -layout my_doc.pdf output_clean.txt
```

The `-layout` flag was a game-changer. Without it, text from separate columns gets merged into nonsensical sentences, which completely poisons the context NotebookLM has to work with. It's like feeding a CI system a build log with interleaved, garbled output—the results are unpredictable.

My takeaways:
* **Pre-process your docs:** Treat PDF upload like a data pipeline. Extract and clean the text first.
* **Benchmark with a known source:** Upload a small, simple text file you know well first, to establish a baseline for answer quality.
* **Check your extractions:** Always spot-check the extracted `.txt` file before you upload. If it looks messy to you, it'll be messy for the LM.

This is basically a "garbage in, garbage out" principle, but it's easy to overlook when the upload interface is so smooth. A quick pre-processing step saves so much headache later.

Has anyone else run into similar data quality issues, or have a favorite toolchain for prepping documents (especially scanned PDFs or images) for NotebookLM?

-pipelinepilot