TriDoc — A Great Offline Document Intelligence

Unified document intelligence with three focused modes—Outline, Importance, and Comprehensive—for faster insights, precise facts, and clear overviews, running fully offline across platforms.

Comment

Inspiration

  • Many users have diverse needs when processing documents: from quick previews to deep understanding to extracting structured data.
  • A clear functional split can improve efficiency and user experience compared to a single “all-in-one” analysis.
  • Inferred from current design:
    • Pain point: a single generic pipeline serves poorly when users need different depths of insight (scan vs. deep dive vs. structured facts).
    • Non-goal: avoid a single “do-everything” button; separation is intentional for clarity and efficiency.
  • Offline-first requirement: fully local inference on a lightweight gpt-oss-20b model enables private, reliable analysis on typical laptops.
  • Product form factor: a cross-platform Electron desktop app with a user-friendly GUI and a mode dropdown (Outline, Importance, Comprehensive).

What it does

Our project provides three complementary document-processing capabilities and a fully offline, cross-platform desktop application. The app parses and summarizes content according to different template types, handles long documents via smart chunking with binary-content filtering, and exports results as JSON across all functions.

1) General Analysis

  • Application scenarios:
    • Comprehensive understanding of a document
    • Exploratory analysis when specific needs are unclear
    • Multi-angle, in-depth analysis of complex documents
    • Broad Q&A and interactive analysis
    • May include structure analysis and key point identification
  • Applicable document types:
    • Research reports, academic papers
    • Long-form articles
    • Business proposals or plans
    • Complex technical documentation
  • Inferred from current design:
    • Output emphasis: breadth and depth of insights, ability to answer broad questions.
  • Implementation context:
    • Delivered through a desktop GUI with mode selection; runs locally on gpt-oss-20b with JSON export.

2) Entity Extraction

  • Application scenarios:
    • Identify and extract specific information (persons, locations, organizations, dates, etc.)
    • Build structured databases
    • Mine targeted information from large volumes of documents
    • Information retrieval and classification
    • Construct knowledge graphs
  • Applicable document types:
    • News articles
    • Contracts
    • Resumes
    • Product manuals
    • Clinical medical records
    • Financial reports (e.g., company names, figures, dates)
  • Inferred from current design:
    • Output emphasis: correctness and completeness of specified entity types as structured fields.
  • Implementation context:
    • Outputs can be exported as JSON like the other functions; processed locally with smart chunking when needed.

3) Document Summary

  • Application scenarios:
    • Quickly grasp a document’s main content
    • Save time on large volumes of documents
    • Provide document previews and executive summaries
    • Content aggregation and condensation
  • Applicable document types:
    • Long reports
    • Academic papers
    • News articles
    • Meeting minutes
    • Legal documents
    • Technical white papers
  • Inferred from current design:
    • Output emphasis: readable summaries that cover key points without unnecessary detail.
  • Implementation context:
    • Executed via the desktop GUI; long inputs are chunked locally, with binary content filtered; results exported as JSON.

Why this breakdown is reasonable

  • Clear functional differences targeting different goals
  • Different depth and breadth:
    • General Analysis = depth + breadth
    • Entity Extraction = precise targeting
    • Document Summary = compression and concision
  • Distinct user scenarios; easier to choose the right tool
  • Efficiency: targeted functions avoid unnecessary comprehensive analysis
  • Better user guidance than a single universal option
  • Aligns with common practices in document processing and NLP
  • Inferred from current design:
    • Value: targeted modes reduce compute and user time when only a summary or entities are needed.
    • Non-goal: sentiment analysis, topic modeling, and other specialized tasks are not default—reserved for future extensions.
  • Operational advantages:
    • Entirely offline on a local gpt-oss-20b model; data stays on-device, suitable for privacy-sensitive environments.
    • Cross-platform Electron app with a user-friendly GUI; packaged for one-click install and optimized for everyday hardware.

Possible improvements or alternatives

  • Task-oriented options:
    • “Quick Browse” → Summary
    • “In-Depth Analysis” → General Analysis
    • “Information Extraction” → Entity Extraction
  • More granular functions:
    • Add options like Topic Analysis, Sentiment Analysis
    • Tailor analyses for specific document types (e.g., contracts, resumes)
  • Adaptive analysis:
    • System auto-selects the best method based on document type and content
  • Composite analysis:
    • Allow combining methods, e.g., “Summary + Entity Extraction”
  • Inferred from current design:
    • UX guidance: offer presets for common combinations (e.g., “Summary + Entities”).
    • Roadmap logic: core three modes → composite workflows → adaptive analysis → specialized modules.

How we built it

  • Electron-based desktop app for Windows, macOS, and Linux with a simple, consistent UI.
  • Local‑only inference using a gpt‑oss‑20b model; no network dependency.
  • Chunking strategy for long texts and binary‑content filtering to keep outputs coherent.
  • A unified JSON export pipeline shared across all modes.
  • Packaging and resource optimizations for smooth installs and everyday hardware.
  • Fallback mechanisms to ensure basic results with messy inputs.

Challenges we ran into

  • Balancing on‑device performance and output quality on modest hardware, even with a 20B model.
  • Designing a UI that clearly exposes three modes without overwhelming users.
  • Managing long‑document context (chunking, stitching, de‑duplication) fully offline.
  • Keeping JSON outputs stable and predictable across varied document types and formats.

Accomplishments that we're proud of

  • Defined a clear, user-centered functional framework covering major document-analysis needs.
  • Inferred from current design:
    • Clarified boundaries and non-goals to reduce scope creep.

What we learned

  • Clear separation of analysis goals improves usability and efficiency.
  • Users benefit from both targeted tools and options to combine methods.
  • Inferred from current design:
    • Success signals:
    • Summary: readability and coverage of key points.
    • Entity Extraction: correctness and completeness for specified entities.
    • General Analysis: breadth and depth of insights and ability to answer broad questions.
    • Trade-off awareness: speed vs. depth, precision vs. coverage are key considerations across modes.

What's next for Untitled

  • Explore adaptive analysis that auto-selects the best pipeline per document.
  • Add specialized modules (e.g., Topic Analysis, Sentiment Analysis, contract/resume-specific analyzers).
  • Support composite workflows like “Summary + Entity Extraction.”
  • Inferred from current design:
    • Milestone ordering: 1) solidify three core modes, 2) add common presets for combinations, 3) experiment with adaptive selection, 4) extend with domain-specific modules.

Built With

Share this project:

Updates