DocuBrowse 0.9.1: Local AI-Powered Document Search Engine

DocuBrowse 0.9.1 provides a fully local, AI-powered solution for transforming unstructured collections of documents into a searchable knowledge base. By combining traditional keyword search with semantic AI embeddings, it allows users to find documents based on meaning rather than just exact text matches, all while keeping data on the local machine to ensure privacy and eliminate API costs.

Hybrid Search Architecture

DocuBrowse employs a dual-search mechanism to maximize retrieval accuracy. By default, it uses a Hybrid Mode that merges results from two distinct scoring systems:

  • Keyword Search: Utilizes SQLite FTS5 (Full-Text Search) with BM25 scoring. It performs fast prefix-matching across titles, authors, subjects, tags, and content snippets.
  • Semantic Search: Uses the nomic-embed-text:latest model via Ollama to generate 768-dimensional vectors. It calculates cosine similarity between the query embedding and a cached, L2-normalized embedding matrix in-process using NumPy for low-latency retrieval.

In Hybrid Mode, the final relevance score is calculated as 0.3 × keyword_score + 0.7 × semantic_score, ensuring that conceptually relevant documents are prioritized while still rewarding exact keyword matches.

AI-Powered Document Synopses

To accelerate document review, DocuBrowse generates "Kindle-style" book-jacket synopses on demand. When a user clicks a document title, the system uses the dolphin3:latest model via Ollama to create a concise summary of the content. To optimize performance, these synopses are cached in the SQLite database after the first generation, preventing redundant AI processing for the same file.

Broad Document Format Support

Version 0.9.1 expands indexing capabilities to include a wide array of file types, ensuring most common office and e-book formats are covered:

  • Office Documents: PDF (via pdfplumber and pypdf), DOCX (python-docx), PPTX (python-pptx), and XLSX (openpyxl).
  • OpenDocument Format (ODF): Added in v0.9.1, the system now indexes .odt, .ods, and .odp files using Python's standard library (zipfile and xml.etree.ElementTree) without requiring external dependencies.
  • E-books: EPUB, MOBI, AZW3, and AZW (via ebooklib and Calibre). Note that DRM-encrypted AZW files are indexed by metadata only.
  • Plain Text: HTML, TXT, and Markdown.

Privacy and Security Hardening

Because DocuBrowse is designed for local personal use, it implements several security layers to prevent unauthorized access and data leaks:

  • PII Protection: A post-ingest scanner detects patterns for Social Security Numbers (SSNs), credit cards, bank routing/account numbers, dates of birth, and passports. Matching documents are removed from the index and permanently blacklisted to prevent accidental exposure.
  • Local-Only Binding: The server binds exclusively to the loopback subnet (localhost/127.0.0.1), rejecting all non-loopback connections at the socket level.
  • Web Hardening: To prevent DNS-rebinding and Cross-Site Request Forgery (CSRF), the system uses a host-header allow-list and requires a per-process X-CSRF-Token for all state-changing API mutations (such as deleting files or changing settings).

System Requirements and Deployment

DocuBrowse is distributed as native packages for Linux (RPM, DEB, tarball), Windows (zip), and macOS (dmg).

Hardware Requirements:

  • Minimum: 8 GB RAM, x86_64 or ARM64 CPU, and 2 GB free disk space. It can run without a GPU, though synopsis generation will be slower.
  • Recommended: 16 GB RAM and 4 GB+ vRAM (NVIDIA or Apple Silicon) for GPU-accelerated embeddings and synopses.

Software Prerequisites:

  • Python 3.9+
  • Ollama (automatically installed by docubrowser start if missing)
  • Calibre (required for MOBI/AZW3/AZW indexing)

Community Insights and Use Cases

Users and developers on Hacker News highlighted the value of a "local-first" approach to document management, contrasting it with cloud-based RAG (Retrieval-Augmented Generation) tools.

"Not a fan of pushing every personal document to someone else's cloud. Nice to see a tool that keeps everything on disk instead."

Some users suggested potential integrations with tools like Paperless-NGX or expanding support to S3-compatible targets and cloud drives like Google Drive or Dropbox to bridge the gap between local indexing and cloud storage.

Sources

Related

  • Project
    linuxrebel/DocuBrowser

    A local-first document search engine that uses AI to provide keyword and semantic search, automated synopses, and deep-linking to specific passages across numerous file formats.

  • Project
    opendatalab/OmniDocBench

    OmniDocBench is a real open‑source benchmark for evaluating document‑parsing AI systems. It provides a 1.6k‑page, multilingual PDF dataset with rich block‑ and span‑level annotations (text, tables, formulas, layout, reading order) and an evaluation suite (Docker/Conda) that computes edit‑distance, BLEU, TEDS, CDM, and detection metrics. Designed for end‑to‑end and module‑wise assessment of OCR, table extraction, formula recognition, and layout detection.

  • Project
    PromtEngineer/localGPT

    A private, on-premise document intelligence platform that uses hybrid search and local LLMs to let users query and summarize their files without data leaving their machine.

  • Project
    clusterzx/paperless-ai

    Paperless‑AI is an open‑source extension for the self‑hosted document manager Paperless‑ngx. It automatically classifies new documents, adds AI‑generated titles/tags, and builds a vector index so you can ask natural‑language questions across your whole archive (RAG chat). It works with any OpenAI‑compatible model—including local Ollama servers—and ships as a Docker image with a simple web UI for configuration and manual processing.

  • Project
    chrisryugj/Docufinder

    A local search engine that indexes the content of PC documents (including HWP, PDF, and images) to enable fast content-based search and optional AI-powered RAG.