sift

Your documents stay yours.

Privacy-first document intelligence you can self-host, with citations on every answer.

Built for rooms that lock the door.

sift is a serious tool for legal, medical, research, and enterprise document work. The product bends around trust, not novelty.

  • Privacy first

    Sensitive PDFs, contracts, and research corpora stay in your environment. sift is built for data that cannot leave the building.

  • Self-host by default

    Run on your machines. Docker Compose today; Kubernetes and air-gap installers on the roadmap. Minimally opinionated about your models.

  • Grounded answers

    Every chat reply carries source citations. Retrieval is hybrid dense plus BM25 with reranking, tied to your documents.

  • Air-gap capable

    Designed to operate without outbound model calls when you need isolation. Prefetch models, run locally, keep the network closed.

What sift does

Point sift at a folder of documents. Get a searchable, chattable, agent-ready knowledge base.

Ingest

PDF, DOCX, PPTX, XLSX, CSV, HTML, and images. Layout-aware extraction with confidence scores and human-in-the-loop review.

Retrieve

Hybrid dense and BM25 retrieval with reranking, grounded in your corpus.

Chat

Streamed answers with source citations, session memory, and verification.

Serve agents

A first-class MCP endpoint plus a REST mirror. Scoped, auditable access for LLM agents.

Vault

Markdown vault with wikilinks, backlinks, tags, and a graph view.

How it works

From a folder of files to a cited answer, in four steps.

  1. 01

    Point at a folder

    Drop documents into a watched path. PDF, Office, HTML, CSV, and images are accepted.

  2. 02

    Parse and review

    Layout-aware extraction produces structured chunks with confidence scores. Review low-confidence spans before they enter the index.

  3. 03

    Retrieve

    Hybrid dense and BM25 search with reranking returns grounded passages from your corpus.

  4. 04

    Chat or serve agents

    Ask in the web UI with citations, or give agents scoped access through the MCP endpoint and REST mirror.

Trust and deployment

sift is built for teams that cannot send sealed contracts or patient records to a third-party SaaS. Run it where the documents already live. Keep network egress optional.

  • Self-hostable on your infrastructure
  • Air-gap capable with local model prefetch
  • Scoped agent access over MCP and REST
  • Auditable retrieval and citation trails
  • Human review for low-confidence parses
  • CLI, web app, and Compose profiles

Status

Pre-alpha, in active development

The repository is intentionally sparse at the top level while the product is being built. Public artifacts land as they stabilize. External contributions are not open yet.

Source on GitHub