Index PDF elements with mixed embedding models

PDFs are rich with both text and visual content, from descriptive paragraphs to illustrations and tables. This example builds an end-to-end flow that parses, embeds, and indexes both, with full traceability to the original page. For a simpler variant that converts whole PDFs to Markdown, see the PDF to Markdown example.

In this example, we split out both text and images, link them back to page metadata, and enable unified semantic search. We’ll use CocoIndex to define the flow, SentenceTransformers for text embeddings, and CLIP for image embeddings, all stored in Qdrant for retrieval.

🔍 What it does

This flow automatically:

• Extracts page texts and images from PDF files
• Skips tiny or redundant images
• Creates consistent thumbnails (up to 512×512)
• Chunks and embeds text with SentenceTransformers (all-MiniLM-L6-v2)
• Embeds images with CLIP (openai/clip-vit-large-patch14)
• Stores both embeddings and metadata (e.g., where the text or image came from) in Qdrant, enabling unified semantic search across text and image modalities

📦 Prerequisite:

Run Qdrant

If you don’t have Qdrant running locally, start it via Docker:

docker run -p 6333:6333 -p 6334:6334 qdrant/qdrant

📁 Input Data

We’ll use a few sample PDFs (board game manuals). Download them into the source_files directory:

./fetch_manual_urls.sh

Or, feel free to drop in any of your own PDFs.

⚙️ Run the flow

Install dependencies:

pip install -e .

Then build your index (sets up tables automatically on first run):

cocoindex update --setup main

Or run in CocoInsight

cocoindex server -ci main

Define the flow

Flow definition

Let’s break down what happens inside the PdfElementsEmbedding flow.

@cocoindex.flow_def(name="PdfElementsEmbedding")
def multi_format_indexing_flow(
    flow_builder: cocoindex.FlowBuilder, data_scope: cocoindex.DataScope
) -> None:
    data_scope["documents"] = flow_builder.add_source(
        cocoindex.sources.LocalFile(
            path="source_files", included_patterns=["*.pdf"], binary=True
        )
    )
    text_output = data_scope.add_collector()
    image_output = data_scope.add_collector()

We define the flow, add a source, and add data collectors.

For flow definition, the decorator:

@cocoindex.flow_def(name="PdfElementsEmbedding")

marks the function as a CocoIndex flow definition, registering it as part of the data indexing system. When executed via CocoIndex runtime, it orchestrates data ingestion, transformation, and collection.
When started, CocoIndex’s runtime executes the flow in either one-time update or live update mode, and incrementally embeds only what changed.

Process each document

Extract PDF documents

We iterate through each document row and run a custom transformation that extracts PDF elements.

with data_scope["documents"].row() as doc:
    doc["pages"] = doc["content"].transform(extract_pdf_elements)

Extract PDF elements

Define a dataclass for structured extraction: we want to extract a list of PdfPage, each of which has a page number, text, and a list of images.

@dataclass
class PdfImage:
    name: str
    data: bytes

@dataclass
class PdfPage:
    page_number: int
    text: str
    images: list[PdfImage]

Next, define a CocoIndex function called extract_pdf_elements that extracts both text and images from a PDF file, returning them as structured, page-wise data objects.

@cocoindex.op.function()
def extract_pdf_elements(content: bytes) -> list[PdfPage]:
    """
    Extract texts and images from a PDF file.
    """
    reader = PdfReader(io.BytesIO(content))
    result = []
    for i, page in enumerate(reader.pages):
        text = page.extract_text()
        images = []
        for image in page.images:
            img = image.image
            if img is None:
                continue
            # Skip very small images.
            if img.width < 16 or img.height < 16:
                continue
            thumbnail = io.BytesIO()
            img.thumbnail(IMG_THUMBNAIL_SIZE)
            img.save(thumbnail, img.format or "PNG")
            images.append(PdfImage(name=image.name, data=thumbnail.getvalue()))
        result.append(PdfPage(page_number=i + 1, text=text, images=images))
    return result

The extract_pdf_elements function reads a PDF file from bytes and extracts both text and images from each page in a structured way. Using pypdf, it parses every page to retrieve text content and any embedded images, skipping empty or very small images to avoid noise.

Each image is resized to a consistent thumbnail size (up to 512×512) and converted into bytes for downstream use. The result is a clean, per-page data structure (PdfPage) that contains the page number, extracted text, and processed images, making it easy to embed and index PDFs for multimodal search.

Process each page

Once we have the pages, we process each page.

1. Chunk the text

This takes each PDF page’s text and splits it into smaller, overlapping chunks.

with doc["pages"].row() as page:
    page["chunks"] = page["text"].transform(
        cocoindex.functions.SplitRecursively(
            custom_languages=[
                cocoindex.functions.CustomLanguageSpec(
                    language_name="text",
                    separators_regex=[
                        r"\n(\s*\n)+",
                        r"[\.!\?]\s+",
                        r"\n",
                        r"\s+",
                    ],
                )
            ]
        ),
        language="text",
        chunk_size=600,
        chunk_overlap=100,
    )

2. Process each chunk

Embed and collect the metadata we need. Each chunk includes its embedding, original text, and references to the filename and page where it originated.

with page["chunks"].row() as chunk:
    chunk["embedding"] = chunk["text"].call(embed_text)
    text_output.collect(
        id=cocoindex.GeneratedField.UUID,
        filename=doc["filename"],
        page=page["page_number"],
        text=chunk["text"],
        embedding=chunk["embedding"],
    )

3. Process each image

We use CLIP to embed the image and collect the data, embedding, and metadata filename, page_number.

with page["images"].row() as image:
    image["embedding"] = image["data"].transform(clip_embed_image)
    image_output.collect(
        id=cocoindex.GeneratedField.UUID,
        filename=doc["filename"],
        page=page["page_number"],
        image_data=image["data"],
        embedding=image["embedding"],
    )

When we collect image outputs, we also want to preserve relevant metadata alongside the embeddings. For each image, we not only collect the image embedding and binary image data, but also metadata such as the original file name and page number. This association makes it possible to trace embedded images back to their source document and page during retrieval or exploration.

Export to Qdrant

Finally, we export the collected data to the target store.

text_output.export(
    "text_embeddings",
    cocoindex.targets.Qdrant(
        connection=qdrant_connection,
        collection_name=QDRANT_COLLECTION_TEXT,
    ),
    primary_key_fields=["id"],
)
image_output.export(
    "image_embeddings",
    cocoindex.targets.Qdrant(
        connection=qdrant_connection,
        collection_name=QDRANT_COLLECTION_IMAGE,
    ),
    primary_key_fields=["id"],
)

🧭 Explore with CocoInsight (free beta)

Use CocoInsight to visually trace your data lineage and debug the flow.

It connects locally with zero data retention.

Start your local server:

cocoindex server -ci main

Then open the UI 👉 https://cocoindex.io/cocoinsight

💡 Why this matters

Traditional document search only scratches the surface: it’s text-only and often brittle across document layouts. This flow gives you multimodal recall, meaning you can:

• Search PDFs by text or image similarity
• Retrieve related figures, diagrams, or captions
• Build next-generation retrieval systems across rich content formats

Compare with ColPali vision model (OCR free)

We also have an example for ColPali.

To compare the two approaches:

In essence, ColPali excels in deep, integrated vision-text understanding, ideal for visually complex documents, while splitting text and images for separate embeddings is more modular but prone to losing spatial and semantic cohesion. The choice depends on your document complexity, precision needs, and resource constraints.

Support us

⭐ Star CocoIndex on GitHub and share with your community if you find it useful!

CocoIndex

Fresh context for long-horizon agents.

Learn the concept → View on GitHub

About the author.

Linghua Jin Cofounder & CEO of CocoIndex · @linghuaj

CEO and cofounder of CocoIndex. Writes about incremental data infrastructure, agents, and the engineering decisions behind the engine.