​
Building a PDF Search Workflow

Pixeltable PDF search works in two phases:

  1. Define your workflow structure (once)
  2. Query your document database (anytime)
1

Install Dependencies

pip install pixeltable tiktoken sentence-transformers

Define Your Workflow

Create table.py:

import pixeltable as pxt
from pixeltable.iterators import DocumentSplitter
from pixeltable.functions.huggingface import sentence_transformer

# Initialize app structure
pxt.drop_dir("pdf_search", force=True)
pxt.create_dir("pdf_search")

# Create documents table
documents_t = pxt.create_table(
    "pdf_search.documents", 
    {"pdf": pxt.Document}
)

# Create chunked view for efficient processing
documents_chunks = pxt.create_view(
    "pdf_search.document_chunks",
    documents_t,
    iterator=DocumentSplitter.create(
        document=documents_t.pdf,
        separators="token_limit",
        limit=300  # Tokens per chunk
    )
)

# Configure embedding model
embed_model = sentence_transformer.using(
    model_id="intfloat/e5-large-v2"
)

# Add search capability
documents_chunks.add_embedding_index(
    column="text",
    string_embed=embed_model
)

Use Your Workflow

Create app.py:

import pixeltable as pxt

# Connect to your tables
documents_t = pxt.get_table("pdf_search.documents")
documents_chunks = pxt.get_table("pdf_search.document_chunks")

# Sample document URLs
DOCUMENT_URL = (
    "https://github.com/pixeltable/pixeltable/raw/release/docs/resources/rag-demo/"
)

document_urls = [
    DOCUMENT_URL + doc for doc in [
        "Argus-Market-Digest-June-2024.pdf",
        "Company-Research-Alphabet.pdf",
        "Zacks-Nvidia-Report.pdf",
    ]
]

# Add documents to database
documents_t.insert({"pdf": url} for url in document_urls)

# Search documents
query = "What are the growth projections for tech companies?"
top_n = 3
sim = documents_chunks.text.similarity(query)
result = (
    documents_chunks.order_by(sim, asc=False)
    .select(documents_chunks.text, sim=sim)
    .collect()
)

# Print results
for i in result:
    print(f"Similarity: {i['sim']:.3f}")
    print(f"Text: {i['text']}\n")

​
What Makes This Different?

Smart Chunking

Token-aware document splitting:

iterator=DocumentSplitter.create(
    document=documents_t.pdf,
    separators="token_limit",
    limit=300
)

Vector Search

Natural language document search:

documents_chunks.add_embedding_index(
    column="text",
    string_embed=embed_model
)

Auto-updating

Self-maintaining document database:

documents_t.insert([{"pdf": new_url}])
# Chunking and embeddings update automatically

​
Workflow Components

​
Advanced Usage

​
Custom Chunking Strategies

Configure different chunking approaches:

# Chunk by paragraphs
chunks_by_para = pxt.create_view(
    "pdf_search.para_chunks",
    documents_t,
    iterator=DocumentSplitter.create(
        document=documents_t.pdf,
        separators="paragraph"
    )
)

# Chunk by fixed size
chunks_by_size = pxt.create_view(
    "pdf_search.size_chunks",
    documents_t,
    iterator=DocumentSplitter.create(
        document=documents_t.pdf,
        separators="fixed",
        size=1000  # characters
    )
)

​
Batch Processing

Process multiple PDFs in batch:

# Bulk document insertion
pdf_urls = [
    "https://example.com/doc1.pdf",
    "https://example.com/doc2.pdf",
    "https://example.com/doc3.pdf"
]

documents_t.insert({"pdf": url} for url in pdf_urls)

​
Advanced Search Functions

Create specialized search functions:

@pxt.query
def search_with_metadata(
    query: str,
    min_similarity: float,
    limit: int
):
    sim = documents_chunks.text.similarity(query)
    return (
        documents_chunks.where(sim >= min_similarity)
        .order_by(sim, asc=False)
        .select(
            documents_chunks.text,
            documents_chunks.pdf_name,
            documents_chunks.page_number,
            similarity=sim
        )
        .limit(limit)
    )