Deep Research Engine

v0.2
Back to App

AI-Powered Deep Research for Engineers

Upload PDFs, paste URLs, link GitHub repos — and get a detailed, cited engineering report with verifiable sources. Not a summary. A real research document with evidence-backed claims.

What Does It Do?

Ingest Sources

Upload PDFs, paste web URLs, or link GitHub repositories. The system extracts, chunks, and indexes all content.

Deep Reports

A 5-step pipeline: Plan → Retrieve → Write → Judge → Refine. Every claim is cited. Insufficient evidence is flagged.

Flashcards

Auto-generate Q&A flashcards from any report. Export as Anki-compatible CSV or JSON for study.

How To Use

1

Add Your Sources

Click 'Attach PDF' to upload research papers, or 'Paste URLs' to add web pages and GitHub repos. Each source gets chunked and indexed.

2

Ask Your Question

Type a specific engineering question. The more specific, the better the report. Example: 'Compare Mamba vs Transformer for real-time inference on edge devices'.

3

Choose Your Mode

'Answer' for quick responses. 'Deep Report' for the full research pipeline with citations. 'Flashcards' to generate study cards from a completed report.

4

Review & Export

Read the report in the Report tab. Check Sources tab for all citations. Export flashcards as Anki CSV. Download the report as Markdown.

Research Pipeline

StepUses AI?What It Does
📋 Planner✅ YesBreaks your question into sub-questions + must-check items (baselines, failure modes, gotchas)
🔍 Retrieval❌ NoBM25 keyword search — finds the top-k most relevant chunks per sub-question from your sources
✍️ Writer✅ YesWrites the full report with [source:chunk] citations. Follows a strict format with trade-off tables
🔍 Judge✅ YesChecks for missing citations, contradictions, shallow sections. Scores quality 0-100%
🔧 Refiner✅ If neededOnly runs if Judge score < 70%. Regenerates flagged sections, then re-judges

Technology Stack

Next.js 15

React frontend framework

FastAPI

Python async backend

OpenRouter

Free LLM models (primary)

Groq

Fast LLM fallback

BM25

Keyword search (pure Python)

SQLite

Local database

PyMuPDF

PDF text extraction

Tavily

Optional web search API

SSE

Real-time streaming events

LLM Strategy: Free-First + Fallback

Primary: OpenRouter — uses openrouter/free which auto-selects the best available free model (Llama 3.3, Gemini Flash, DeepSeek, etc). Zero cost.

Fallback: Groq — if OpenRouter returns 429 (rate limit) or 5xx (server error), the system automatically retries once, then falls back to Groq with llama-3.3-70b-versatile.

This gives you free-first behavior while keeping the platform reliable when free pools are saturated.