retriever

Retriever

Retriever is an open source document intelligence plugin for Claude Code. It turns local folders, productions, mailbox exports, and other supported document collections into a workspace you can search, browse, review, enrich, analyze, and export with Claude. Users work with Retriever through Claude Code: install it once, open Claude Code in the target workspace, and use plain-English requests or /retriever:* commands to ingest, inspect, review, process, and export. It fits legal teams, investigations, compliance, diligence, and other document-heavy workflows that need local-first search, review, analysis, and handoff.

What Retriever Does

1. Build a document workspace Turn a local folder, processed production, PST/MBOX archive, or other supported source into a Retriever workspace that keeps search state, previews, datasets, processing runs, and exports together.
2. Search, review, enrich, and analyze documents Find key documents and communications, preview files, filter and page through results, run first-pass review and triage, add metadata, run OCR or extraction, and keep the workspace context intact between turns.
3. Export results and handoff artifacts Once the scope is right, export a shareable table, preview bundle, or archive for handoff, QA, downstream loading, or another tool.

Typical uses include first-pass review and hot-doc triage, contract review, mailbox and production analysis, Bates lookup, OCR or extraction over a selected set, and CSV or archive export for handoff.

Project docs:

• PRIVACY.md
• ARCHITECTURE.md
• TESTING.md
• CONTRIBUTING.md
• docs/browse-reference.md

Privacy

IMPORTANT: IF YOU HANDLE SENSITIVE, CLIENT, OR INTERNAL DOCUMENTS, TURN OFF Help improve Claude IN Settings -> Privacy BEFORE USING RETRIEVER. Anthropic says that when this setting is off, new Claude chats and coding sessions are not used for future model training, although flagged conversations may still be used for trust and safety purposes. Review the Anthropic privacy policy and confirm it satisfies your practice, client, regulatory, and organizational requirements before using Retriever with sensitive material.

Zero Data Retention (ZDR) is available for Claude Code on Claude for Enterprise. Anthropic says ZDR is enabled per organization and covers Claude Code inference on Claude for Enterprise. See Zero data retention for the current scope, limitations, and enablement details.

How to Install

Open Claude Code and paste this. Claude does the rest.

Install Retriever: run if [ -d ~/.claude/skills/retriever/.git ]; then git -C ~/.claude/skills/retriever pull --ff-only origin main; else git clone --single-branch --depth 1 https://github.com/sdemyanov/retriever.git ~/.claude/skills/retriever; fi. Then run cd ~/.claude/skills/retriever && ./setup.

To test Retriever with public sample data, open Claude Code and paste this:

Set up Retriever sample data: run if [ -d ~/retriever-data-public/.git ]; then git -C ~/retriever-data-public pull --ff-only origin main; else git clone --single-branch --depth 1 https://github.com/sdemyanov/retriever-data-public.git ~/retriever-data-public; fi.

Then open Claude Code in ~/retriever-data-public and run Ingest this folder.

Quick Start

1. Open Claude Code in the local folder for one document collection, matter, inbox export, or dataset.
2. Ask Claude Code to ingest that folder with Retriever.
3. Ask for the next job in plain English or use /retriever:* commands to search, review, analyze, enrich, and export.

Example prompts:

• Ingest this folder
• Find documents mentioning indemnification
• Review the workspace for hot docs
• Show emails from Alice in 2023
• Extract counterparties from the current contracts
• Export the current results

Suggested first session

If you are trying Retriever for the first time, after opening Claude Code in the target folder, a good starting point is:

• Check the workspace status for this folder
• Ingest this folder
• Show me the first results for <your first keyword>
• Only show emails if that helps
• Add dataset name to the visible columns if you want more context
• Save this scope as first-pass if you want to reuse it
• Export the current results if you want a handoff artifact

That path exercises the setup, browse, narrowing, display, persistence, and export surfaces that most users rely on first.

Supported inputs

Supported file types

• PDFs
• DOCX
• TXT, Markdown, CSV, JSON, HTML, ICS, and many code/config text formats
• EML and MSG, including one level of extracted attachment children
• RTF
• XLS and XLSX
• PPTX
• common image formats as preview-only documents

Supported archive and export sources