why my document is not OCR executed

[doriansong]

I identified a scan using the following configuration:

ingestor = (
    Ingestor(client=client)
    .files(file_arr)
    .extract(   
        document_type="pdf",          
        extract_text=True,
        # extract_tables=True,
        # extract_charts=True,
        extract_images=True,
        # extract_infographics=True,
        # paddle_output_format="markdown",
        # extract_method="nemoretriever_parse", # Slower, but maximally accurate, especially for PDFs with pages that are scanned images
        text_depth="page"
    )
    .split( 
        chunk_size=1024,
        chunk_overlap=50,
        params={"splite_source_types":['PDF']}
    )
    .embed(  
        endpoint_url="http://172.16.200.113:9997/v1",
        model_name="bge-m3",
    )
  
)

print the results using print(ingest_json_results_to_blob(results[0]))

The output shows that the pipe line mistook my scan for a picture
Is it my configuration problem？

Sample scanned document：

An example of the output results：

image_caption:[The image provided is a drawing of a large snake-like creature that appears to be a kind of serpent or dragon in an exaggerated, whimsical style. The creature has several pairs of feet and is winding through various objects and people.

Key elements include:
- A long serpentine body with multiple segments
- A cartoonish, almost comical appearance 
- Tentacle-like appendages emanating from its head and body 
- A trio of duck-like heads   
- It is shown interacting with townsfolk in a small village setting

The style is lighthearted and cartoonish, depicting this fantastical creature as more comical than menacing despite its intimidating size. The overall tone seems playful, presenting the creature's intrusion on the village life as an amusing encounter rather than a frightening threat.]

[Advait251206]

What you’re seeing is usually not a bug in NeMo Retriever, but a consequence of how the pipeline decides whether to run OCR vs treating a page as an image.

---

1. Key point: OCR is not automatically triggered just because a PDF is scanned

In your config:

.extract(
    document_type="pdf",
    extract_text=True,
    extract_images=True,
    text_depth="page"
)

You enabled:

• text extraction
• image extraction

But you did not explicitly enable a forced OCR / layout parsing path.

So the pipeline typically does:

“Try native PDF text extraction first → if no text layer → fallback decisions depend on extractor mode”

If the document is a scanned PDF, but:

• pages are detected as image objects
• or text layer is missing/empty
• or layout classifier decides “image-heavy page”

👉 then it may route it into image understanding instead of OCR

---

2. Why your output shows image_caption instead of OCR text

This line is the key:

image_caption: [...]

That means:

• the pipeline treated the page as an image input
• it sent it to a vision captioning model
• NOT to OCR extraction

So instead of:

OCR text → extracted characters

you got:

image → captioning model → semantic description

---

3. Most likely cause in your config

❗ Missing explicit OCR / parsing strategy

You commented out the key option:

# extract_method="nemoretriever_parse"

This is important because:

• default extractor often prioritizes layout + vision classification
• OCR is only guaranteed in specific parsing modes (like nemoretriever_parse)

---

4. Why scanned PDFs get misclassified

Even if it looks like a scan, the system may classify it as:

Case A: Image-only page

→ goes to vision captioning

Case B: Mixed layout page

→ partial OCR + image fallback

Case C: Low-confidence text detection

→ OCR skipped to avoid garbage extraction

So your output suggests:

the page was classified as "image-heavy", not "text-recoverable"

---

5. How to fix it

✔ Option 1 (recommended): force NemoRetriever parser

Uncomment:

extract_method="nemoretriever_parse"

This is specifically designed for:

• scanned PDFs
• layout-heavy documents
• OCR + structure extraction

---

✔ Option 2: explicitly enable OCR path (if supported by your build)

Depending on version, you may need:

extract_text=True
ocr=True   # if available in your version

or equivalent OCR backend configuration.

---

✔ Option 3: disable image caption fallback (if you want strict OCR)

If your goal is only text extraction, you may want:

• disable extract_images=True
• or disable vision fallback pipeline

Otherwise the system prefers “best semantic understanding” over raw OCR.

---

6. Why text_depth="page" matters

With:

text_depth="page"

you are telling the system:

“aggregate at page level”

So even if OCR runs internally, you may not see granular token-level output unless the parser is explicitly OCR-driven.

---

7. Summary

Your issue is likely caused by:

The pipeline falling back to vision-based image captioning instead of OCR

because:

• OCR parsing mode is not explicitly enabled
• default extraction strategy prefers image understanding for scanned pages
• image_caption output confirms vision route, not OCR route

---

8. Direct answer

✔ Not your fault in document
✔ Not a broken pipeline
❗ Configuration is missing explicit OCR enforcement (nemoretriever_parse or OCR mode)

---

If you want, I can help you rewrite the ingestion pipeline so it:

• always uses OCR for scanned PDFs
• never falls back to image captioning
• and still preserves layout structure

---

If this answer helped or pointed you in the right direction, I'd appreciate it if you could mark it as the accepted answer so it's easier for others with the same issue to find.

Also, if you found my contribution useful, I'd appreciate it if you could check out my GitHub profile, follow me, and star any repositories you find interesting.

GitHub: https://github.com/Advait251206