langchain-fastocr 0.1.1

A LangChain document loader that reads scanned PDFs and images with FastOCR.

Website · Docs · Get an API key

Quickstart

pip install langchain-fastocr
export FASTOCR_API_KEY="your_api_key"

from langchain_fastocr import FastOCRLoader

documents = FastOCRLoader("scanned_contract.pdf").load()
print(documents[0].page_content)

Python 3.10 or newer. The distribution is langchain-fastocr and the import is langchain_fastocr.

What you get back

FastOCRLoader returns one Document per file.

If the account runs out of pages partway through a file, you still get the text of the pages that were processed, with status set to partial and a warning in the log. A document that fails raises the error from fastocr-sdk with its message and code.

Languages and files

• 100+ languages, detected automatically. No language setting is needed.
• PDF files up to 1 GiB, and PNG, JPG, JPEG, GIF, WebP and BMP images up to 20 MiB. An image counts as one page.

Get an API key and pages

Any signed-in FastOCR account can create an API key and start right away.

1. Sign in and create an API key in the developer console. You accept the API terms with your first key.
2. Set the key as FASTOCR_API_KEY, or pass api_key= to the loader.
3. When you need more than the free pages, buy a Page Pack on the pricing page. Packs work on the website and in the API.

Every account gets 50 free pages a month, shared with PDFs on the website. An image sent to the API is one page.

Examples

RAG over scanned PDFs

from langchain_core.vectorstores import InMemoryVectorStore
from langchain_fastocr import FastOCRLoader
from langchain_openai import OpenAIEmbeddings
from langchain_text_splitters import RecursiveCharacterTextSplitter

documents = FastOCRLoader("scanned_contract.pdf").load()
chunks = RecursiveCharacterTextSplitter(chunk_size=1000, chunk_overlap=150).split_documents(documents)

store = InMemoryVectorStore.from_documents(chunks, OpenAIEmbeddings())
for hit in store.as_retriever(search_kwargs={"k": 3}).invoke("When can the contract be terminated?"):
    print(hit.metadata["file_name"], hit.page_content[:200])

Each chunk keeps its file's metadata, so answers can cite the source file. This example also needs pip install langchain-openai langchain-text-splitters numpy and an OpenAI key; any other embeddings class works the same way.

A folder of scans

from pathlib import Path

from langchain_fastocr import FastOCRLoader

scans = sorted(p for p in Path("scans").iterdir() if p.suffix.lower() in {".pdf", ".png", ".jpg", ".jpeg"})
for document in FastOCRLoader(scans).lazy_load():
    print(document.metadata["file_name"], document.metadata["pages_billed"])

lazy_load() yields each document as soon as its file is finished. The loader runs one file at a time.

Searchable PDFs for an archive

from langchain_fastocr import FastOCRLoader

documents = FastOCRLoader(["invoice-001.pdf", "invoice-002.pdf"], output="pdf").load()
for document in documents:
    print(document.metadata["searchable_pdf_path"])

Each searchable PDF is saved beside its input, as invoice-001.searchable.pdf. An existing file with that name is replaced. Index the text and keep the PDF.

Options

Support and license

Questions: support@fastocr.org. Bugs and requests: GitHub issues. MIT license; see LICENSE.