AXIOM AI - Trust Every Answer, Proof Before Prose

AXIOM AI is a sophisticated Research Copilot and Knowledge Management platform designed to ingest, process, and query complex research documents. The platform is built around a robust Retrieval-Augmented Generation (RAG) architecture, ensuring that every AI answer is strictly grounded in the provided source material.

Note: This README is maintained automatically and will be updated iteratively after every prompt by the AI assistant.

---

🚀 Key Features

• Document Knowledge Graph: Upload PDFs and automatically extract, chunk, and embed them into a secure vector database.
• Strict Contextual RAG: Vector searches are securely filtered by user and active document using precise metadata definitions (document_id, user_id).
• Research Workspace (Split-Screen): A premium 60/40 UI allowing side-by-side conversational analysis and inline PDF viewing via secure signed URLs.
• Structured Rendering: AI responses are dynamically formatted with markdown-to-Tailwind parsers, providing beautifully structured lists, code blocks, and citations without visual clutter.
• Quick Summarization: Generate instant, structured overviews of entire documents directly from the chat interface.
• Audio Synthesis: High-quality Text-to-Speech (TTS) capabilities built directly into individual copilot messages.
• Research Paper Comparison: Select 2–5 papers and generate structured side-by-side comparisons covering methodology, architecture, performance, and key insights.
• Startup Idea Generator: Transform research insights into actionable startup concepts with structured idea cards (problem, solution, target users, tech stack).
• SaaS Identity Management: Complete profile lifecycle, authentication, and JWT authorization powered by Supabase.

---

🛠 Tech Stack

Frontend

• Framework: React / Vite
• Styling: TailwindCSS & Custom Glassmorphism CSS Modules (glass-panel)
• Animation: GSAP (GreenSock) for cinematic user experiences
• Routing: React Router DOM
• Markdown Rendering: react-markdown mapped directly to layout design tokens

Backend

• Framework: FastAPI (Python)
• Authentication & Core DB: Supabase (PostgreSQL / GoTrue JWT / Storage)
• Vector Database: Qdrant (Primary local vector store implementation)
• RAG Framework: LlamaIndex RouterQueryEngine for semantic search and reliable answer generation
• Document Processing: pypdf for text parsing and extraction

---

⚙️ Architecture Workflow

1. Ingestion: Uploaded PDFs are stored in Supabase Storage. Text is parsed, chunked semantically, and embedded.
2. Indexing: Embeddings are pushed to Qdrant alongside meticulously injected metadata (e.g., user_id, document_id).
3. Retrieval: When a query is made, LlamaIndex applies ExactMatchFilter using the required authentication token and precise document ID to ensure heavily sandboxed retrieval.
4. Generation: The designated LLM reviews the chunks to form a structured answer, ultimately returning explicit source chunk citations directly to the React frontend.

---

💻 Local Setup & Development

Prerequisites

• Node.js (v18+)
• Python 3.10+
• Qdrant Server running locally on port 6333
• Supabase Project (URL and Anon/JWT Keys)

Environment Variables (.env)

You will need to configure your respective .env files in both frontend and backend directories with the active Supabase endpoints, Qdrant host parameters, and LLM API keys.

Running the Services

1. Start the Backend API:

cd backend
python -m venv venv
source venv/Scripts/activate  # (Windows)
pip install -r requirements.txt
uvicorn main:app --reload --port 8000

2. Start the Frontend Application:

cd axiomai-ui
npm install
npm run dev

Visit http://localhost:5173 to interact with the AXIOM AI dashboard.

---

✨ New Feature: Research Paper Comparison

Overview

Compare multiple research papers and generate structured, LLM-powered insights without re-processing any documents.

How It Works

1. Navigate to COMPARE in the dashboard navbar
2. Select 2–5 uploaded & indexed papers
3. Click COMPARE — the system retrieves stored vectors for each paper
4. A structured comparison prompt is sent to the LLM
5. Results render as a side-by-side table + insight cards

API Endpoint

POST /api/compare
Body: { "paper_ids": ["id1", "id2"] }

Output Structure

• Methodology comparison
• Dataset / Benchmark comparison
• Architecture differences
• Performance metrics comparison
• Strengths & Weaknesses per paper
• Key Insights (cross-paper observations)

Tech Used

• LlamaIndex (vector retrieval with metadata filters)
• Qdrant (read-only vector search per document_id)
• Ollama / LangChain (LLM generation)
• Supabase (paper metadata + auth)

Notes

• Uses already-indexed data — no re-ingestion required
• Multi-user isolated: only the logged-in user's papers are accessible
• Maximum 5 papers per comparison to stay within LLM context limits

---

💡 New Feature: Startup Idea Generator

Overview

Generate innovative, practical startup ideas based on your research paper insights.

How It Works

1. Navigate to IDEAS in the dashboard navbar
2. Select 1–5 indexed research papers
3. Click GENERATE IDEAS — the system extracts key themes and insights
4. LLM generates 3–5 structured startup concepts
5. Results render as detailed idea cards

API Endpoint

POST /api/generate-ideas
Body: { "paper_ids": ["id1", "id2"] }

Output per Idea

• Title — Catchy startup concept name
• Description — Elevator pitch
• Problem — What real-world problem it addresses
• Solution — How the product works
• Target Users — Who benefits
• Tech Stack — Suggested technologies

Tech Used

• LlamaIndex (vector retrieval with metadata filters)
• Qdrant (read-only vector search per document_id)
• Ollama / LangChain (LLM generation)
• Supabase (paper metadata + auth)

---

🐛 Recent Bug Fixes

Issue	Root Cause	Fix
QA returns "Answer not found" on valid documents	Documents ingested before metadata injection lack document_id in vectors, causing ExactMatchFilter to return 0 results	Added fallback unfiltered retrieval in router_agent.py when metadata-filtered search yields empty
Summary renders raw HTML tags (<p>, <strong>)	LLM was outputting HTML instead of Markdown	Rewrote SUMMARY_PROMPT to enforce clean Markdown; added stripHtml() safety layer in StructuredMessageRenderer.jsx

---

📝 Changelog

• v1.4 — AI Startup Idea Generator (research-to-startup pipeline)
• v1.3 — Research Paper Comparison (multi-paper structured analysis), Document deletion (Qdrant + Storage + DB cleanup)
• v1.2 — Structured Message Renderer, Summary Quick-Action, QA fallback retrieval, HTML→Markdown sanitization
• v1.1 — Split-Screen Research Workspace, Document metadata injection, Vector filter grounding
• v1.0 — Initial RAG pipeline, Auth system, Document upload, Chat interface