How liweiphys/layra works
Project Type and Purpose
• Layra is a full-stack, retrieval-augmented generation (RAG) platform with a Next.js/TailwindCSS frontend and a FastAPI-based asynchronous backend.
• It ingests PDF documents as images, generates visual embeddings, stores metadata and vectors, and exposes a chat-style QA interface that preserves layout.
File-Structure Insights
• Top-level directories:
• app/ ← FastAPI backend, organized into api, core, db, models, schemas, rag, utils, framework
• web/ ← Next.js 15 + TypeScript frontend
• docker/ ← Docker-Compose definitions for Milvus, Redis, MongoDB, Kafka, MinIO
• model_server.py ← standalone ColBERT embedding service
• Key config files: .env, alembic.ini, docker-compose.yml, gunicorn_config.py, next.config.ts
• Asynchronous/event-driven: Kafka consumers/producers, SSE endpoints, background embedding jobs.
README Architecture Sections
• Query flow: embedding → vector retrieval → answer generation
• Upload flow: PDF → image pages → ColQwen2.5 embeddings → Milvus + metadata stores
• Tech stack: FastAPI, Kafka, Redis, MySQL, MongoDB, MinIO, Milvus, Next.js
Main Components and Interactions
a. Frontend (Next.js)
• Renders UI pages (chat, knowledge base, sign-in)
• Calls REST/API (chatApi, configApi, kbApi)
• Maintains auth/chat/config state via Zustand
b. Backend (FastAPI)
• HTTP endpoints (app/api/endpoints/*.py) for auth, chat, SSE, config
• Core modules for config, logging, security
• Schemas/models define request/response and DB ORM
• RAG services (app/rag) invoke embedding service and LLM
• Framework layer (app/framework/app_framework.py) sets up middleware, routers
c. Embedding Service (model_server.py)
• Hosts ColQwen2.5 visual embedding model
• Receives page images, returns dense vectors
d. Data Stores
• Milvus: vector similarity search
• MySQL (via SQLAlchemy/Alembic): relational metadata (users, configs)
• MongoDB: document metadata, conversation history
• MinIO: raw PDF files and page images
• Redis: caching, session/state management
e. Message Broker / Background Jobs
• Kafka:
• Producer: when PDF is uploaded, produce “new-document” event
• Consumer: long-running job to parse PDF → images → embedding → store
f. External LLMs
• Qwen2.5-VL (and future GPT-4o, Claude) invoked via llm_service
g. Clients/Users
• Browser users interact via web UI over HTTPS
• Optionally any OpenAI-compatible client via API
Architectural Patterns and Principles
• Async-first, event-driven pipeline (FastAPI + Kafka)
• Microservice-style separation: embedding service vs. core API
• RAG pattern: separation of retrieval (Milvus) and generation (LLM)
• Clean layering: API → services → data access → external integrations