Implement SBERT Vector Embeddings and Stateful Context for Guard Agent

[ttudii]

Overview

This PR delivers the final, feature-complete implementation of the G·U·A·R·D Authentication System.

It finalizes the transition from syntactic matching to a Semantic Engine using sentence-transformers (SBERT). The system is now fully Dockerized, operationally Stateful (handling ambiguous inputs via short-term memory), and documented for deployment.

Closes Issue: #4 (Experimental Semantic Auth)

---

Key Changes Implemented

1. Semantic Engine & Zero-Knowledge Storage

• Integrated SBERT (all-MiniLM-L6-v2) for 384-dimensional vector embeddings.
• Zero-Knowledge Architecture: Raw passphrases are vectorized immediately upon enrollment and discarded. Only the embeddings are stored.
• Ambiguity Logic: Implemented the "Contextual Challenge" flow (Score 0.65–0.80 triggers a clarification request).

2. Database Architecture (SQLite)

Moved from flat files to a structured SQLite database (guard_secure.db) with three tables:

• users: Credentials and lock status.
• phrases: Serialized JSON vector embeddings.
• auth_context: Short-term memory table to handle multi-turn ambiguous verifications.

3. DevOps & Deployment

• Added Dockerfile for the FastAPI backend.
• Added docker-compose.yml to orchestrate the backend and Ngrok tunnel automatically.
• Secured configuration via .env file injection.

4. Documentation

• Completely rewrote README.md to reflect the new architecture.
• Added Frontend Guide.
• Updated API Reference for all endpoints (/enroll, /verify, /user/{id}, /update).

---

Checklist

• Backend Logic (SBERT + FastAPI)
• Database Implementation (SQLite + Context Table)
• Dockerization (Compose + Ngrok)
• Zero-Knowledge Text Handling verified
• Documentation & README updated