LexiLingo

The AI English Tutor That Actually Understands You

GraphCAG · Real-Time Voice · Knowledge Graph · CEFR Assessment

---

Most language apps give you the same lesson regardless of who you are. LexiLingo builds a live knowledge graph of your gaps, diagnoses your errors in real time, and generates personalized explanations — not templates.

Get Started · Architecture · Why GraphCAG? · Features

---

What Makes LexiLingo Different

Traditional Apps	LexiLingo
Fixed curriculum for all learners	Knowledge Graph maps YOUR concept gaps
Generic AI explanations	GraphCAG grounds every response in your mastery state
Slow responses from heavy RAG	Redis-cached context → sub-second grounding
One language model for everything	Smart Router picks the right model per task
No pronunciation depth	HuBERT phoneme analysis with Vietnamese-specific patterns
Fixed review intervals	SuperMemo-2 spaced repetition, EF-adjusted per concept

---

Why GraphCAG over RAG?

RAG retrieves from documents. GraphCAG grounds responses in a live knowledge graph of the learner.

	Traditional RAG	LexiLingo GraphCAG
Context source	Static document chunks	Live KuzuDB knowledge graph + Redis learner cache
Personalization	None — same docs for everyone	Per-user mastery scores, error history, CEFR level
Retrieval latency	Vector search on every turn	Pre-cached learner profile, graph hop expansion
Curriculum awareness	Zero — just similarity search	Prerequisite chains: "Past Simple → Past Perfect → Reported Speech"
Error diagnosis	Not possible	Dedicated Diagnose node maps errors to KG concept IDs
Routing logic	Single retrieval path	Conditional edges: low-confidence → clarify, A1/A2 errors → Vietnamese bridge
State between turns	Stateless or basic memory	Full GraphCAGState TypedDict persisted across turns

Result: Responses feel like a real tutor who remembers everything about you — not a chatbot that reads from a wiki.

---

Architecture

Flutter App (iOS · Android · Web)
         │
    API Gateway (Kong / Nginx)
         │
    ┌────┴────┬──────────────┐
    ▼         ▼              ▼
Backend    AI Service     Admin
(FastAPI)  (FastAPI)    (React+Vite)
    │         │
PostgreSQL  Redis + KuzuDB

GraphCAG Pipeline (LangGraph StateGraph)

The core of LexiLingo's intelligence — a 7-node stateful graph that replaces naive RAG:

INPUT ──▶ KG_EXPAND ──▶ DIAGNOSE
               (KuzuDB)     (Error→KG IDs)
                                │
              ┌─────────────────┼──────────────┐
              ▼                 ▼              ▼
         ASK_CLARIFY      VIETNAMESE        RETRIEVE
         (conf < 0.5)    (A1/A2 bridge)  (KG + Cache)
              │                 │              │
              └─────────────────┴──────────────┘
                                │
                            GENERATE
                            (LLM + CAG)
                                │
                       ┌────────┴────────┐
                       ▼                 ▼
                      TTS              END

Each node is a pure async def node(state) -> dict updating a shared GraphCAGState. LangGraph handles conditional routing, retries, and state checkpointing.

Smart Model Router

User Input
    │
    ▼
Complexity Analysis
    │
    ├─ Simple greeting   →  gemma2:2b  (local, ~3s)
    ├─ Grammar / tech    →  Gemini     (cloud, ~2s)
    └─ Deep analysis     →  qwen3:4b   (local, ~20s)

Fallback chain: OpenRouter → Gemini → Ollama

Dual-Stream Real-Time Voice

Three async streams running concurrently:

LISTENING  ──▶  THINKING  ──▶  SPEAKING
(VAD + STT)    (GraphCAG)    (Chunked TTS)
     │                              │
     ◄────── INTERRUPTION ──────────┘

User can interrupt the AI mid-sentence. TTS stops immediately, STT resumes.

---

Features

AI & Learning Engine

• GraphCAG Pipeline — Knowledge Graph + Cache-Augmented Generation via LangGraph
• CEFR Assessment — Auto-scoring A1→C2 across grammar (40%), vocabulary (30%), fluency (30%)
• HuBERT Pronunciation — IPA phoneme analysis with Vietnamese-specific error patterns (θ→t, ʃ→s, ð→d)
• SM-2 Spaced Repetition — EF-adjusted review intervals, overdue priority queue
• Content Auto-Generation — Exercises targeting your exact error patterns, not generic templates

Voice Pipeline

• Real-time VAD → Streaming STT (Whisper) → GraphCAG → Chunked TTS (Piper)
• Interruption handling — speak over the AI at any time
• Partial transcript updates during speech

Knowledge Graph (KuzuDB)

• Grammar concepts from A1 basics to C2 advanced
• Prerequisite chain traversal for next-best-concept recommendations
• Per-user mastery scores updated live after each interaction

Cross-Platform App

• Flutter 3.24+ — iOS, Android, Web from a single codebase
• Offline-first with SQLite local storage
• 7 locales: English, Vietnamese, Japanese, Korean, Chinese, French, Spanish

---

Technology Stack

Layer	Technology
AI Orchestration	LangGraph StateGraph
LLM (Cloud)	Google Gemini API
LLM (Local)	Ollama (gemma2, qwen3) · Qwen2.5 LoRA fine-tune
Pronunciation	HuBERT-large (Facebook)
STT / TTS	Whisper · Piper
Knowledge Graph	KuzuDB
Backend	FastAPI · PostgreSQL 14+ · Redis · SQLAlchemy
Frontend	Flutter 3.24+ · Provider · Dio
Infrastructure	Docker Compose · Kong Gateway

---

Quick Start

Prerequisites: Python 3.11+, Flutter 3.24+, PostgreSQL 14+, Docker (optional)

git clone https://github.com/InfinityZero3000/LexiLingo.git
cd LexiLingo
bash scripts/start-all.sh

# Docker
docker-compose up -d

---

Contributing

Contributions are welcome — especially on the AI pipeline, Flutter UI, and benchmark tooling.

See CONTRIBUTING.md for guidelines.

---

If LexiLingo gave you ideas, a helps others find it.

Documentation · Report Issue · Discussions

MIT License · Built with by InfinityZero3000