Question-led retrieval
Search a document in plain language and receive the best-matching passages instead of dumping the entire file into context.
Why it exists
A 300-page PDF should not become a 300-page prompt.
DocSift pays the conversion cost once, stores a content-addressed local index, and retrieves the few passages that answer the question. The saving comes from retrieval: less context, less noise, and source locations that remain attached.
~21k
tokens in the measured 34-page source document
~4k
tokens returned for one measured question
0
cloud APIs, accounts, or telemetry required
Built for agent workflows
Useful context, with its provenance intact.
Page-aware citations
Every result keeps its document name, page number, and section heading so the source can be checked quickly.
One document interface
PDFs use Docling; Word, Excel, PowerPoint, HTML, CSV, and EPUB use MarkItDown behind the same CLI and MCP surface.
Local by design
Files, extracted text, indexes, and searches stay on your machine. There are no accounts, hosted services, or analytics.
How it works
From file to answer in three honest steps.
Convert
Layout-aware engines turn the source into clean Markdown while preserving useful document structure.
Index
DocSift creates token-budgeted chunks and identifies the document by its content, even if the file later moves.
Retrieve
The CLI or MCP server returns only matching passages, ready for Claude, Codex, Copilot, Cursor, or another local client.
Put your documents within reach.
Python 3.11+ · choose the conversion engines you need · runs locally over CLI, MCP, or HTTP API.
pip install "docsift[all]"