English · 简体中文

Drop a folder. Ask anything. Keep it local.

DocFlow is a local-first document Q&A and personal knowledge workspace. Point it at a folder of PDFs, Markdown, DOCX, TXT, code-like files, or optional images. Ask questions in the browser, inspect the cited sources, and save useful answers back into your local knowledge loop.

The screenshots in this README are captured from the bundled demo library, not a personal vault.

git clone https://github.com/lingengyuan/docflow.git && cd docflow
docker compose -f docker-compose.image.yml up

Open http://localhost:8000, choose the demo-library card, and click 导入示例资料 to try a small local library.

The browser UI is Chinese-first today. The language button switches common navigation and status labels while fuller localization continues.

For real answers, run a local model server such as Ollama, LM Studio, or another OpenAI-compatible local endpoint, then select it in Settings. The Docker image starts DocFlow and Qdrant; model weights are still managed by the local model tool you choose.

Expected first-run cost: the public image path currently uses about 0.5 GB for the base app image on the validation machine, plus the local model weights you choose. A typical 7B Ollama model is usually 4-5 GB. Image understanding and Apple Silicon MLX support are optional installs.

• Ask questions across PDFs, Markdown, DOCX, TXT, code-like files, and optional image content.
• Review source snippets, cited chunks, source trails, related files, topics, knowledge cards, and active concepts.
• Save useful answers as notes and connect notes back to their source material.
• Keep metadata in SQLite, vectors in Qdrant, and local-model traffic on your machine when a local backend is selected.
• Verify behavior with tests, browser acceptance, retrieval evaluation, parsing evaluation, faithfulness checks, large-library checks, and an offline network check.

flowchart LR
    A[Folder] --> B[Parse + Chunk]
    B --> C[(SQLite metadata)]
    B --> D[(Qdrant vectors)]
    C --> E[Retrieve + Rerank]
    D --> E
    E --> F[Local model]
    F --> G[Cited answer]

DocFlow ships with zero telemetry, zero analytics, zero automatic error reporting, and zero product-analytics document upload. Your documents, SQLite metadata, Qdrant vectors, backups, and indexes stay on your machine unless you explicitly enable an external feature.

docflow doctor --offline

This command checks the covered local startup, ingest, query, model-status, and source-preview paths for unexpected outbound connections.

External activity is opt-in and bounded: webpage import, model downloads, and cloud model backends only run when you configure or trigger them.

These numbers are release and regression evidence, not broad public leaderboard claims. The public-domain regression is separate from the source-filtered internal retrieval cases, which are useful for project regression but not external benchmark claims.

See Evaluation and Status for raw scope, commands, and current limitations.

DocFlow generates config.yaml from config.example.yaml on first run. Configure watched folders, supported extensions, SQLite path, Qdrant connection, embedding model, local model backend, privacy settings, and answer-quality thresholds there.

python -m venv .venv && source .venv/bin/activate
pip install -e .
docker compose up -d qdrant
docflow demo --create-only
scripts/run_ci.sh
docflow doctor --offline

For source development, use docker compose up --build. Public image startup uses docker-compose.image.yml with ghcr.io/lingengyuan/docflow:edge, while versioned image tags are produced for releases.

Evaluation and release verification commands live in Evaluation and Release.

• main.py - command entry point.
• src/api/ - browser app and HTTP API.
• src/ingest/ - parsing, chunking, indexing, and storage.
• src/query/ - retrieval, reranking, citations, and answer generation.
• frontend/ - browser workspace.
• docs/ - public documentation.
• eval/ - committed retrieval and parsing evaluation inputs.

Features · Architecture · Privacy · Threat Model · Model Licenses · CLI · Development · Evaluation · Release · Status · ADRs · Roadmap

Read CONTRIBUTING.md. Keep changes focused, run tests before opening a PR, and keep the normal browser UI free of command-line or developer-only wording.

MIT. See LICENSE.