GitHub - Hundred-Trillion/L88-Full

L88 — Agentic RAG Knowledge Engine

100% Local · Zero Cloud · Fully Private

L88 is a locally-hosted agentic RAG system for researchers. Upload PDFs, ask questions, and get cited answers — powered by a self-correcting LangGraph pipeline running on your own hardware. Multi-user, multi-session, with role-based access control.

Architecture

┌────────────────────────────────────────────────────────────────────┐
│                      Frontend (React + Vite)                       │
│   Sidebar (Sessions)  │  ChatPanel (Messages)  │  RightPanel       │
│   Role Switcher       │  Web Toggle · Upload   │  Docs · Notes     │
│   Session CRUD        │  Markdown Rendering    │  Members           │
├────────────────────────────────────────────────────────────────────┤
│                    Backend (FastAPI + SQLModel)                     │
│   Auth (JWT)  │  Sessions  │  Documents  │  Members  │  Scratchpad │
├────────────────────────────────────────────────────────────────────┤
│                  Intelligence (LangGraph + Ollama)                  │
│   Router → Analyzer → Rewriter → Retriever → Generator             │
│                              ↑          Self-Evaluator ↓           │
│                              └──── retry loop ─────────┘           │
└────────────────────────────────────────────────────────────────────┘

Layer	Stack
Frontend	React 19, Vite 6, Tailwind CSS 4, Framer Motion, Lucide
Backend	FastAPI, SQLModel (SQLite), Pydantic, JWT
Intelligence	LangGraph (agentic orchestration), Ollama (local LLM)
Retrieval	FAISS (vector store), BGE (embeddings + reranking)
LLM	Qwen2.5-AWQ (4-bit quantized, GPU-optimized)

Agentic RAG Pipeline

L88 doesn't just retrieve — it reasons. The backend runs a cyclic graph that self-corrects:

Router — Routes to RAG (docs selected) or general chat (no docs, LLM trained knowledge)
Analyzer & Rewriter — Decomposes complex queries into search-optimized sub-tasks
Retrieval — Queries session documents via FAISS with BGE embeddings + reranking
Generator — Produces structured answers with context evaluation and inline citations
Self-Evaluator — Validates answers against evidence; retries with rewritten queries if gaps found

When no documents are uploaded, L88 falls back to the LLM's trained knowledge for general conversation.

Features

Session Workspaces — Isolated research sessions with their own documents and chat history
Per-Session Collaboration — Add users with role-based access (Admin / Chat / Read Only)
Document Management — Upload PDFs, toggle selection, delete — per session
Scratchpad — Per-session research notes with autosave and .txt download
Web Mode — Toggle internet-augmented retrieval
General Chat — Works without docs — the LLM answers from trained knowledge
100% Offline — Local LLM, local vector store, no external API calls

Roles & Permissions

Capability	Admin	Chat	Read Only
Send messages	✅	✅	❌
Upload / delete documents	✅	❌	❌
Toggle document selection	✅	✅	❌
Manage session members	✅	❌	❌
Edit scratchpad	✅	❌	❌
View everything	✅	✅	✅

Getting Started

Prerequisites

GPU: NVIDIA with ≥8 GB VRAM
RAM: 16 GB+ recommended
Python: 3.10+
Node.js: 18+
Ollama: ollama.com

1. Clone

git clone https://github.com/Hundred-Trillion/L88-Full.git
cd L88-Full

2. Backend

# Install dependencies
cd l88_backend
python -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt

# Terminal 1: Start Ollama
ollama serve

# Terminal 2: Start backend (from project root, not l88_backend/)
cd ~/path/to/L88-Full
uvicorn l88_backend.main:app --reload --port 8000

3. Frontend

cd l88_frontend
npm install
npm run dev

Open http://localhost:8888

Default Login

Username: admin
Password: admin123

Project Structure

L88-Full/
├── l88_backend/
│   ├── main.py                # FastAPI entry point
│   ├── config.py              # Model & storage config
│   ├── database.py            # SQLModel + user seeding
│   ├── routers/
│   │   ├── auth.py            # JWT authentication
│   │   ├── sessions.py        # Session CRUD
│   │   ├── chat.py            # Chat + RAG trigger
│   │   ├── documents.py       # Document management
│   │   ├── members.py         # Member management
│   │   ├── scratchpad.py      # Session notes
│   │   └── admin/             # Admin endpoints
│   ├── services/              # Business logic
│   ├── graph/                 # LangGraph agentic pipeline
│   │   ├── nodes/
│   │   │   ├── router.py      # Intent routing
│   │   │   ├── query_analyzer.py
│   │   │   ├── query_rewriter.py
│   │   │   ├── retrieval.py   # FAISS search
│   │   │   ├── generator.py   # LLM generation
│   │   │   └── self_evaluator.py
│   │   ├── edges.py           # Conditional routing
│   │   └── graph.py           # Graph assembly
│   ├── ingestion/             # PDF parsing, chunking, embedding
│   └── retrieval/             # FAISS vector store
├── l88_frontend/
│   ├── src/
│   │   ├── App.tsx            # Root + state management
│   │   ├── components/
│   │   │   ├── LoginPage.tsx  # Auth screen
│   │   │   ├── Sidebar.tsx    # Sessions + role switcher
│   │   │   ├── ChatPanel.tsx  # Chat + input + web toggle
│   │   │   └── RightPanel.tsx # Docs + scratchpad + members
│   │   ├── services/api.ts    # Backend API client
│   │   ├── context/           # Auth + Theme providers
│   │   └── types.ts           # TypeScript types
│   └── vite.config.ts
└── README.md

License

Proprietary — L88 Laboratories / Hundred Trillion.