gizmosql-pdf-loader

Load PDF files into a GizmoSQL server over Arrow Flight SQL / ADBC so that SQL clients (and GizmoSQL-connected MCP servers) can search the text of product manuals and download the original files.

For every PDF the loader stores:

Works with plain DuckDB catalogs and with DuckLake catalogs (no primary-key / unique constraints are used; the loader de-duplicates on SHA-256 itself).

Install

python3 -m venv .venv && . .venv/bin/activate
pip install gizmosql-pdf-loader          # or: pip install dist/gizmosql_pdf_loader-*.whl

Dependencies: adbc-driver-gizmosql (>= 2.0.13), pyarrow, pymupdf, click, python-dotenv.

OCR of image-only pages needs no extra install: the Tesseract engine ships inside the PyMuPDF wheel, and the only external piece, the trained language data (eng.traineddata, about 4 MB), is downloaded on first use into ~/.cache/gizmosql-pdf-loader/tessdata from the tessdata_fast repository. An installed Tesseract's language folder (or --tessdata / TESSDATA_PREFIX) is used instead when present; --no-ocr-download forbids the download. Without language data, image-only pages are loaded with empty text and a warning.

Configure

Create a .env in the working directory (or export the variables):

GIZMOSQL_HOSTNAME=gizmosql.example.com
GIZMOSQL_PORT=443
GIZMOSQL_USERNAME=me@example.com
GIZMOSQL_PASSWORD=...
GIZMOSQL_CATALOG=my_lake          # target catalog (must already be attached on the server)
GIZMOSQL_SCHEMA=pdf_docs          # target schema (created if missing); default: pdf_docs
# optional
GIZMOSQL_USE_TLS=true             # false -> gizmosql://host:port?transport=tcp
GIZMOSQL_TLS_SKIP_VERIFY=false
GIZMOSQL_MAX_MESSAGE_SIZE_BYTES=67108864   # client-side gRPC limit (driver default is 16 MiB)

Always set a catalog: GizmoSQL sessions default to the in-memory memory catalog, and the loader refuses to load into it.

Use

# Load every *.pdf in a directory (or pass explicit file paths)
gizmosql-pdf-loader load --source-dir source_pdf_files

# Inspect/extract without connecting to the server
gizmosql-pdf-loader load --dry-run --source-dir source_pdf_files

# What is loaded?
gizmosql-pdf-loader list

# Search page text: substring match, or BM25-ranked via the full-text index
gizmosql-pdf-loader search "hydraulic pump" --limit 5
gizmosql-pdf-loader search "hydraulic pump relief valves" --ranked --limit 5

# (Re)build the full-text index (also done automatically at the end of `load`)
gizmosql-pdf-loader index

Output of a real load (six equipment manuals, 641 MB in total, into a DuckLake catalog backed by S3):

status            size chunks  pages w/text   ocr       chars    secs  file
loaded      73,588,282      9    457    457     0     561,719    46.7  operator-manual-1.pdf
loaded      74,815,458      9    424    424     0     556,911    52.5  operator-manual-2.pdf
loaded     113,837,480     14   1688   1688     0   1,217,423    29.8  parts-manual-1.pdf
loaded       6,783,584      1     84     84     0      60,788    10.0  parts-manual-2.pdf
loaded     359,084,382     43   3004      6     0          26   105.6  shop-manual-1.pdf
loaded      12,740,509      2     94      0     0           0    12.5  shop-manual-2.pdf

Useful load options (see gizmosql-pdf-loader load --help for all):

Documents are identified by a UUID derived from the file's SHA-256, so the same bytes always map to the same document_id; re-running the loader is idempotent.

Use as a Python library

Everything the CLI does is a plain function call; the top-level package exports the pieces you need (GizmoSQLSettings, LoadOptions, OcrSettings, ensure_schema, load_pdf, plan_fts, build_fts_index, inspect_pdf, extract_pages, ...). examples/load_with_python.py is the runnable version of this:

import sys
from pathlib import Path

from dotenv import load_dotenv

from gizmosql_pdf_loader import (
    GizmoSQLSettings,
    LoadOptions,
    OcrSettings,
    build_fts_index,
    ensure_schema,
    load_pdf,
    plan_fts,
)

load_dotenv(dotenv_path=".env")

CATALOG = "my_lake"
SCHEMA = "pdf_docs"
source_dir = Path(sys.argv[1] if len(sys.argv) > 1 else "source_pdf_files")

# GizmoSQLSettings(hostname=..., port=..., username=..., password=...) also works without env vars.
settings = GizmoSQLSettings.from_env()

with settings.connect(catalog=CATALOG) as conn:
    ensure_schema(conn, catalog=CATALOG, schema=SCHEMA)  # idempotent: tables, views, macros, migrations

    options = LoadOptions(
        catalog=CATALOG,
        schema=SCHEMA,
        chunk_size=8 * 1024 * 1024,
        replace=False,  # True reloads files whose SHA-256 is already present
        ocr=OcrSettings(enabled=True, language="eng"),
    )
    for pdf_path in sorted(source_dir.glob("*.pdf")):
        result = load_pdf(conn, path=pdf_path, options=options)
        r = result
        print(f"{r.status:8} {r.file_name}: {r.page_count} pages, {r.total_text_chars:,} chars")

    # Full-text index: built in place on a DuckDB-file catalog, skipped on DuckLake unless a
    # DuckDB catalog is named to hold a copy of the page text.
    plan = plan_fts(conn, catalog=CATALOG, schema=SCHEMA, fts_catalog="memory")
    print(f"fts: {plan.mode} ({plan.reason})")
    pages_indexed = build_fts_index(conn, plan)

    with conn.cursor() as cur:
        if pages_indexed:
            cur.execute(
                f"SELECT file_name, page_number, score "
                f"FROM {CATALOG}.{SCHEMA}.search_pages_ranked(?, top_k := 3)",
                parameters=["relief valve"],
            )
        else:
            cur.execute(
                f"SELECT file_name, page_number, snippet FROM {CATALOG}.{SCHEMA}.search_pages(?) LIMIT 3",
                parameters=["relief valve"],
            )
        for row in cur.fetchall():
            print(row)

load_pdf returns a LoadResult (status is "loaded" or "skipped", plus chunk/page/text counts and the document_id). inspect_pdf(path) and extract_pages(path, ocr=...) work without any server connection if you only want the metadata or the text.

Query from SQL

-- Ranked full-text search (BM25, English stemming) via the fts index
SELECT file_name, page_number, score, snippet
  FROM my_lake.pdf_docs.search_pages_ranked('hydraulic pump relief valves', top_k := 10);

-- Substring search with a snippet (case-insensitive, no index needed)
SELECT file_name, page_number, snippet
  FROM my_lake.pdf_docs.search_pages('relief valve')
 LIMIT 10;

-- Same thing by hand
SELECT d.file_name, p.page_number, p.page_text
  FROM my_lake.pdf_docs.document_pages p
  JOIN my_lake.pdf_docs.documents d USING (document_id)
 WHERE p.page_text ILIKE '%relief valve%';

-- Metadata
SELECT file_name, mime_type, file_size_bytes, page_count, pages_with_text, title, producer, pdf_created_at
  FROM my_lake.pdf_docs.documents;

-- Fetch a file's chunks in order (each row is one chunk, so any client message limit works)
SELECT chunk_index, content
  FROM my_lake.pdf_docs.document_chunks
 WHERE document_id = (SELECT document_id FROM my_lake.pdf_docs.documents
                       WHERE file_name = 'parts-manual-2.pdf')
 ORDER BY chunk_index;

Full-text index

search_pages_ranked is backed by DuckDB's fts extension (BM25 scoring, English stemmer). The loader looks up the target catalog's type in duckdb_databases() and logs what it decides:

Loading the fts extension needs the GizmoSQL admin role; querying an existing index does not. The index is a snapshot: load refreshes it after loading new documents, and gizmosql-pdf-loader index rebuilds it on demand. If the copy catalog is the server's in-memory memory catalog, the copy and index vanish on a server restart, so run index again afterwards (or point --fts-catalog at an attached DuckDB file).

# DuckLake target: opt in to a copy-based index in the shared in-memory catalog
gizmosql-pdf-loader index --fts-catalog memory

Download a file from Python

examples/download_document.py streams the chunks in order and verifies the SHA-256, so it works with the driver's default 16 MiB message limit:

python examples/download_document.py "parts-manual-2.pdf" /tmp/parts-manual-2.pdf

The core of it:

cur.execute(
    f"SELECT content FROM {qs}.document_chunks WHERE document_id = ? AND chunk_index = ?",
    parameters=[document_id, chunk_index],
)
(content,) = cur.fetchone()
out.write(content)

Hidden layers ("secure" PDFs)

Some vendor products (e.g. HyperGEAR PDFLib "secure documents") do not encrypt a PDF; they put the real content on an optional-content layer whose default view state is OFF and rely on viewer JavaScript to switch it on after checking a license window. Outside that viewer every page looks blank, and OCR finds nothing.

The loader records such layers in documents.hidden_layer_names and, by default, leaves them alone. With --reveal-hidden-layers it writes a temporary copy with those layers switched ON, extracts text from the copy, deletes it, and sets documents.hidden_layers_revealed = true. The stored file bytes are always the original. Use this only for documents you are entitled to read outside the vendor's viewer.

How the gRPC message limit is handled

• Uploads use cursor.adbc_ingest with a RecordBatchReader, one record batch per chunk, so each Flight message carries at most one chunk (8 MiB by default) plus a little overhead.
• The 16 MiB figure is the client driver's default (adbc.flight.sql.client_option.with_max_msg_size), not a server limit: a GizmoSQL server accepted and served back a single 256 MiB message in testing. The loader raises its own limit to 64 MiB (GIZMOSQL_MAX_MESSAGE_SIZE_BYTES overrides it); other clients can keep their defaults because they only ever read one chunk per row.
• Page text goes up 200 pages per batch.
• There is intentionally no server-side view that reassembles a whole file. DuckDB needs roughly ten times the file size in memory to concatenate large BLOBs; reassembling the 359 MB file above pushed the PoC server past its 4.4 GiB limit and restarted it. Reassemble on the client (examples/download_document.py) instead.

Development

pip install -e ".[dev]"
ruff check src tests examples
pytest -q                      # the integration test runs only when GIZMOSQL_HOSTNAME is set
python -m build                # wheel + sdist in dist/

Releases: bump version in pyproject.toml, move the [Unreleased] notes in CHANGELOG.md into a version section, commit, then git push origin main vX.Y.Z. CI runs the tests on Python 3.10 through 3.14, builds the wheel and sdist, creates a GitHub Release with that CHANGELOG section as the notes and the artifacts attached, and publishes the package to PyPI via trusted publishing (the pypi GitHub environment; no token secrets).