Research Assistant

Overview

Research Assistant is a local-first, offline research paper assistant designed to process scientific publication PDFs and automatically generate structured deliverables. Without requiring external API keys or remote LLM calls, the assistant generates full-paper extractive summaries (section-aware), simplified versions, podcast-style scripts with optional Text-to-Speech (TTS) WAV audio, formatted PowerPoint presentation decks, and video scripts along with evaluation metrics.

Problem

Processing complex research papers often forces users to rely on cloud-hosted LLM APIs, raising privacy concerns and requiring recurring subscription or token costs. Furthermore, simple text models often produce “abstract-only” summaries by failing to analyze section distributions or mathematical sentence scoring. Building a fully offline, private research assistant requires combining high-fidelity local PDF extraction engines, OCR fallbacks for scanned documents, and computationally efficient extractive summarization algorithms (such as TF-IDF combined with MMR diversity optimization) that execute quickly on consumer hardware.

Approach

The tool processes papers via a modular, local-first pipeline:

1. Document Parsing & Extraction: Reads embedded text pages. If standard text extraction fails, it triggers an OCR fallback using pdf2image (Poppler) and Tesseract OCR to extract clean text.
2. Text Preprocessing: Normalizes inputs by removing hyphenations, stripping out running headers/footers, cleaning citations, and fixing common PDF parsing artifacts.
3. Section-Aware Summarization: Computes sentence importance using TF-IDF and Maximal Marginal Relevance (MMR) or TextRank. It applies a section-aware allocation algorithm to guarantee that sections like methodology and results are represented, preventing the summary from being dominated solely by the abstract.
4. Deliverable Generation: Formats the extracted highlights into simplified summaries, writes structured podcast scripts, converts them to audio via offline Text-to-Speech, creates a structured .pptx deck via python-pptx, and outputs sentence scoring logs and metadata.

Architecture

graph TD
    A[Research PDF] --> B[PDF Extraction Engine]
    B -->|Has Selectable Text| C[Text Preprocessor]
    B -->|Scanned Image Fallback| D[Tesseract OCR]
    D --> C
    C --> E[Section-Aware Allocator]
    E --> F[TF-IDF + MMR / TextRank]
    F --> G[Processed Highlights]
    G --> H[Deliverables Generator]
    H --> I[Summary & Simplified Text]
    H --> J[python-pptx Deck]
    H --> K[TTS Audio Engine - WAV]
    H --> L[Sentence Scores CSV]

Results

The offline research assistant runs entirely locally, ensuring complete data privacy for proprietary papers and drafts. By using lightweight statistical scoring (TF-IDF + MMR) instead of neural network inference, full paper analysis and deck generation are completed within seconds, rather than minutes. The section-aware allocator ensures comprehensive summarization of the methodology and discussion, improving key detail coverage compared to standard top-n sentence selectors.

Lessons Learned

1. OCR Dependencies on Windows: Implementing OCR fallbacks requires external system binaries (Tesseract and Poppler). Providing fallback alerts is essential for smooth user experiences when dependencies are missing.
2. Audio Post-Processing: Direct Text-to-Speech outputs can have uneven pacing. Leveraging pydub and FFmpeg to stitch segments and adjust silences significantly enhances the listenability of the podcast-style audio.
3. Preventing Abstract Bias: Standard extractive summaries are heavily biased toward the abstract and introduction. Forcing a stratified sampling across sections guarantees that critical findings in results and methodologies are captured.