sirdas (Python)

Local document parsing for AI agents and machine learning. PDF (including scans), Word, Excel, PowerPoint, OpenDocument, EPUB, HTML, CSV, email and images become Markdown, fields cited with page and position, tables, form fields, cited chunks and training datasets. Nothing is uploaded: it runs on your machine.

pip install sirdas          # needs Node.js 20+ (the engine is shared with the CLI and MCP server)

import sirdas

doc = sirdas.read("factura.pdf")                 # one call does it all
doc["document_type"]                             # "factura"
{f["name"]: f["value"] for f in doc["fields"]}   # número, CUFE, NIT, IVA, total… each with page, evidence and bbox
doc["next_steps"]                                # what to do next

sirdas.convert("informe.docx")["markdown"]
sirdas.split_documents("lote-escaneado.pdf", output_dir="separados/")
sirdas.forms("formulario.pdf")                   # filled fields and checkboxes
sirdas.dataset(["a.pdf", "b.docx"], "ds/", format="alpaca")   # + Hugging Face dataset card

LangChain and LlamaIndex:

from sirdas.integrations import SirdasLoader, SirdasReader
docs = SirdasLoader("contrato.pdf").load()              # langchain Documents, stable ids
nodes = SirdasReader().load_data("historia.pdf")        # llama_index Documents

Personal data is anonymized by default (anonymize=False to keep it). Password-protected PDFs raise PasswordRequired; pass password=.

Docs: https://sirdas.app/docs/agents/python.md