中文版

This directory contains the deployment, startup, and image build scripts for EdgeCraftRAG.

1.Script Overview

The main scripts in this directory are:

  • quick_start.sh: recommended one-click deployment script for new users, with automatic setup and interactive guidance

  • bootstrap.sh: non-interactive deployment orchestrator that can be used directly or invoked by quick_start.sh

  • model_download.sh: model preparation helper (supports vllm / ov, optional model_id and model_path arguments)

  • run_ov_baremetal.sh: OpenVINO bare-metal startup script

  • run_ov_container.sh: OpenVINO container startup script

  • run_vllm_baremetal.sh: vLLM bare-metal startup script

  • run_vllm_container.sh: vLLM container startup script

  • run_ovms_baremetal.sh: OVMS bare-metal startup script

  • run_ovms_container.sh: OVMS container startup script

  • build_images.sh: container image build script

Deployment methods:

Method

Description

Requirements

Milvus Support

baremetal

Start services as Python processes

Python 3.10+

No (in-memory only)

container

Start services in Docker containers

Docker / Docker Compose

Yes (enabled by default)

Note: If you need Milvus, use the container deployment method.

2.Quick Deployment Script (New Users)

2.1 One-Command Quick Deployment

Run this from the EdgeCraftRAG root directory:

./tools/quick_start.sh

The script behaves as follows by default:

  • runs in non-interactive mode

  • uses OpenVINO as the default inference backend

  • if INFERENCE_BACKEND is not set, the script resolves it to openvino

  • uses baremetal as the default deployment method when DEPLOYMENT_METHOD is not set

In the default bare-metal flow, the script automatically:

  • creates and activates EdgeCraftRAG/ecrag_venv if it does not exist

  • validates the Python version (3.10+ required, 3.10/3.11 recommended)

  • checks and installs required Python packages

  • checks and installs npm for baremetal UI startup when needed

  • validates Intel GPU driver/runtime and auto-installs missing packages on apt-based Linux

  • checks and auto-downloads missing models (embedding, reranker, OpenVINO LLM)

  • writes a deployment environment snapshot to workspace/bootstrap.env before invoking bootstrap.sh

  • calls bootstrap.sh to start services

For vLLM deployments and container deployment method, the script also validates Docker and Docker Compose before deployment. On Ubuntu 24.04, if Docker or Docker Compose is missing, the script attempts automatic installation and starts/enables Docker service.

To skip model verification/download when models are already prepared locally:

./tools/quick_start.sh --skip-model-check

Equivalent environment variable:

export SKIP_MODEL_CHECK=1
./tools/quick_start.sh

Intel GPU driver/runtime validation can be skipped when needed:

./tools/quick_start.sh --skip-gpu-driver-check

Equivalent environment variables:

export SKIP_INTEL_GPU_DRIVER_CHECK=1
# Or keep validation but disable auto-install:
export AUTO_INSTALL_INTEL_GPU_DRIVER=0
./tools/quick_start.sh

To disable automatic npm installation during baremetal preparation:

export AUTO_INSTALL_NPM=0
./tools/quick_start.sh

After startup succeeds, the terminal prints a UI access URL such as:

UI access URL: http://${HOST_IP}:8082

Note: If you set DEPLOYMENT_METHOD=container in advance, the script skips venv and pip checks and continues with container deployment.

You can override defaults with environment variables:

export INFERENCE_BACKEND=openvino
export MODEL_PATH="${PWD}/workspace/models"
export DOC_PATH="${PWD}/workspace"
export TMPFILE_PATH="${PWD}/workspace"
export LLM_MODEL="Qwen/Qwen3-8B"
export HOST_IP="$(hostname -I | awk '{print $1}')"

./tools/quick_start.sh

Select the backend with INFERENCE_BACKEND:

# OpenVINO (default)
./tools/quick_start.sh

# vLLM_A770
export INFERENCE_BACKEND=vllm_a770
./tools/quick_start.sh

# vLLM_B60
export INFERENCE_BACKEND=vllm_b60
./tools/quick_start.sh

# OVMS
export INFERENCE_BACKEND=ovms
export OVMS_SOURCE_MODEL=OpenVINO/Qwen3-8B-int4-ov
export OVMS_MODEL_NAME=OpenVINO/Qwen3-8B-int4-ov
export OVMS_TARGET_DEVICE=GPU.0
./tools/quick_start.sh

For OVMS deployments, the tooling now exports the compose-facing variables directly. The most commonly overridden ones are OVMS_SOURCE_MODEL, OVMS_MODEL_NAME, OVMS_TARGET_DEVICE, OVMS_TOOL_PARSER, and OVMS_MAX_NUM_BATCHED_TOKENS.

Important OVMS behavior:

  • OVMS_SOURCE_MODEL keeps your original model ID as-is (for example Qwen/Qwen3-8B).

  • quick_start.sh and bootstrap.sh both persist OVMS variables into workspace/bootstrap.env for reuse.

  • You can replay the exact OVMS configuration with source workspace/bootstrap.env && ./tools/bootstrap.sh.

Compatibility note: the legacy environment variable COMPOSE_PROFILES is still accepted, but new configurations should use INFERENCE_BACKEND.

Supported INFERENCE_BACKEND values:

  • openvino

  • vllm_a770

  • vllm_b60

  • ovms

2.2 Interactive Mode

./tools/quick_start.sh -i

Interactive mode is suitable for first-time deployment or when you are not sure about the parameters. After you run ./tools/quick_start.sh -i, the script prompts step by step and generates the deployment configuration for the current run.

The interactive flow typically includes:

  • choosing the inference backend: OpenVINO / vLLM_A770 / vLLM_B60 / OVMS

  • choosing the deployment method: baremetal / container

  • configuring key parameters: HOST_IP, MODEL_PATH, DOC_PATH, TMPFILE_PATH, LLM_MODEL

  • confirming the configuration and starting deployment, then printing the access URL at the end

Interactive mode is recommended when:

  • this is your first installation and you are not familiar with the environment variables or defaults

  • you need to switch quickly between different hardware targets or inference backends

  • you want to review parameters before deployment to reduce configuration mistakes

Example:

cd EdgeCraftRAG
./tools/quick_start.sh -i

2.3 Common Interactive Input Examples

The following examples show common inputs during the interactive flow. Actual prompt text may vary slightly based on the script.

Example A: OpenVINO + baremetal (single-machine quick experience)

Inference backend: OpenVINO
Deployment method: baremetal
HOST_IP: 192.168.1.20
MODEL_PATH: /home/scale/edgeai/applications.edge.ai.rag/EdgeCraftRAG/workspace/models
DOC_PATH: /home/scale/edgeai/applications.edge.ai.rag/EdgeCraftRAG/workspace
TMPFILE_PATH: /home/scale/edgeai/applications.edge.ai.rag/EdgeCraftRAG/workspace
LLM_MODEL: Qwen/Qwen3-8B
Confirm deployment: y

Example B: vLLM_B60 + container (Milvus required)

Inference backend: vLLM_B60
Deployment method: container
HOST_IP: 192.168.1.20
MODEL_PATH: /home/scale/edgeai/applications.edge.ai.rag/EdgeCraftRAG/workspace/models
DOC_PATH: /home/scale/edgeai/applications.edge.ai.rag/EdgeCraftRAG/workspace
TMPFILE_PATH: /home/scale/edgeai/applications.edge.ai.rag/EdgeCraftRAG/workspace
LLM_MODEL: Qwen/Qwen3-8B
Confirm deployment: y

Example D: OVMS + container

Inference backend: OVMS
Deployment method: container
HOST_IP: 192.168.1.20
MODEL_PATH: /home/scale/edgeai/applications.edge.ai.rag/EdgeCraftRAG/workspace/models
DOC_PATH: /home/scale/edgeai/applications.edge.ai.rag/EdgeCraftRAG/workspace
TMPFILE_PATH: /home/scale/edgeai/applications.edge.ai.rag/EdgeCraftRAG/workspace
LLM_MODEL: Qwen/Qwen3-8B
Confirm deployment: y

Notes:

  • for a remote server, set HOST_IP to an address reachable by the client machine

  • if you need persistent vector retrieval data, use the container deployment method

  • if the device is Intel Arc A770, prefer the vllm_a770 configuration

Cleanup:

./tools/quick_start.sh cleanup

3.Startup Scripts

3.1 bootstrap.sh (Non-Interactive Orchestration)

Run with environment variables defined in advance:

export INFERENCE_BACKEND=openvino
export DEPLOYMENT_METHOD=baremetal
./tools/bootstrap.sh

Use defaults (openvino + baremetal):

./tools/bootstrap.sh

Configuration reuse:

  • quick_start.sh writes workspace/bootstrap.env before real deployment starts.

  • bootstrap.sh also persists configuration for reuse.

  • For OVMS, this includes OVMS_SOURCE_MODEL, OVMS_MODEL_NAME, OVMS_TARGET_DEVICE, OVMS_TOOL_PARSER, and related OVMS_* runtime variables.

source workspace/bootstrap.env
./tools/bootstrap.sh

3.3 model_download.sh (Model Preparation)

Basic usage:

./tools/model_download.sh <mode> [model_id] [model_path]

Modes:

  • vllm: prepare embedding/reranker OpenVINO models + vLLM LLM model

  • ov: prepare embedding/reranker OpenVINO models + OpenVINO INT4 LLM model

Optional arguments:

  • model_id: overrides LLM_MODEL for current run

  • model_path: overrides MODEL_PATH for current run

Examples:

./tools/model_download.sh vllm
./tools/model_download.sh ov Qwen/Qwen3-8B /data/models

Environment behavior:

  • if a virtual environment is already active, it is reused

  • otherwise, the script creates/activates ecrag_venv automatically (same style as quick_start.sh)

  • missing python3-venv / pip prerequisites are installed automatically when supported by the system package manager

3.2 Direct Startup Scripts

You can also call the following scripts directly based on inference backend and deployment method:

  • OpenVINO baremetal: ./tools/run_ov_baremetal.sh

  • OpenVINO container: ./tools/run_ov_container.sh

  • vLLM baremetal: ./tools/run_vllm_baremetal.sh

  • vLLM container: ./tools/run_vllm_container.sh

  • OVMS baremetal: ./tools/run_ovms_baremetal.sh

  • OVMS container: ./tools/run_ovms_container.sh

This is useful when you already know your parameters and want to skip the one-click onboarding flow.

4.Container Image Build Script

Build all images:

./tools/build_images.sh

Build by component:

./tools/build_images.sh mega
./tools/build_images.sh server
./tools/build_images.sh ui
./tools/build_images.sh all

For complete deployment guidance, see ../docs/Advanced_Setup.md.