image-ocr-local-aipc（图像 ocr 本地 aipc）

Image OCR (Windows · GLM-OCR · llama.cpp Vulkan)

Model: ggml-org/GLM-OCR-GGUF (Q8_0, HuggingFace / hf-mirror) Inference: llama-cli (llama.cpp Vulkan prebuilt) SKILL_VERSION: v1.0

Directory Structure (auto-created or user-specified)

<OCR_DIR>\                        ← auto-selected drive or user-specified (e.g. C:\image-ocr or D:\image-ocr)
├── llama.cpp\                    ← llama-cli.exe and related binaries
└── models\
    └── GLM-OCR-GGUF\
        ├── GLM-OCR-Q8_0.gguf        ← main model (~950 MB)
        └── mmproj-GLM-OCR-Q8_0.gguf ← vision projection layer (~484 MB, required)

Dependencies: Model files (GLM-OCR-Q8_0.gguf, mmproj-GLM-OCR-Q8_0.gguf) are downloaded via Python's huggingface_hub (hf download) or modelscope. If Python is not installed, Step 2 will automatically install Miniforge (recommended — lightweight, includes conda/pip, no admin rights required).

⚠️ AI Assistant Instructions

1. Execute one command at a time; wait for output before proceeding.
2. Stop immediately on error; refer to the Troubleshooting table at the end.
3. Wrap all paths in double quotes.
4. <OCR_DIR> is the absolute working directory path, determined after Pre-flight.
5. Single goal: Recognize image content and return text results.

Execution flow (do not skip steps):

Pre-flight: Check working dir + llama.cpp + models      → STATUS values
Step 1:     Install / update llama.cpp (only if MISSING) → LLAMA_OK
Step 2:     Download models (only if MISSING)            → MODEL_OK
Step 3:     Process recognition result + output          → Return result

Progress reporting: Announce each step before starting, e.g.: 🔍 Pre-flight: Checking environment…

Pre-flight: Check Environment

🔍 Pre-flight: Checking working directory, llama.cpp, and model files…

Locate Working Directory

# ── Fix encoding for non-ASCII paths (required at the start of every PowerShell script) ──
chcp 65001 | Out-Null
[Console]::OutputEncoding = [System.Text.Encoding]::UTF8
$OutputEncoding = [System.Text.Encoding]::UTF8

# ── Optional: if you already have a path, fill it in; leave blank to auto-select drive ──
$customOcrDir = ""   # e.g. "C:\image-ocr" or "D:\image-ocr"
# ──────────────────────────────────────────────────────────────────────────────────────────

if ($customOcrDir -and (Test-Path (Split-Path $customOcrDir))) {
    $OCR_DIR = $customOcrDir
    New-Item -ItemType Directory -Force -Path $OCR_DIR | Out-Null
    Write-Host "OCR_DIR=$OCR_DIR (user-specified)"
} else {
    $best = Get-PSDrive -PSProvider FileSystem |
        Where-Object { $_.Free -gt 0 } |
        Sort-Object Free -Descending |
        Select-Object -First 1
    $OCR_DIR = Join-Path "$($best.Root)" "image-ocr"
    New-Item -ItemType Directory -Force -Path $OCR_DIR | Out-Null
    Write-Host "OCR_DIR=$OCR_DIR (auto-selected drive: $($best.Name))"
}
$env:OCR_DIR = $OCR_DIR

Success criteria: Output contains a line with OCR_DIR=. Record the path and substitute <OCR_DIR> in subsequent steps.

Check llama.cpp

$llamaDir = "<OCR_DIR>\llama.cpp"
$cliExe   = "$llamaDir\llama-cli.exe"

if (Test-Path $cliExe) {
    $ver = & $cliExe --version 2>&1
    if ($ver -match "version:\s*(\d+)") {
        $build = [int]$Matches[1]
        if ($build -ge 8400) {
            Write-Host "OK: llama.cpp build $build >= b8400, skip Step 1"
            Write-Host "LLAMA_STATUS=READY"
        } else {
            Write-Host "WARN: llama.cpp build $build < b8400, upgrade required"
            Write-Host "LLAMA_STATUS=OUTDATED"
        }
    }
} else {
    Write-Host "ERROR: llama-cli.exe not found"
    Write-Host "LLAMA_STATUS=MISSING"
    Write-Host "   Checked path: $llamaDir"
}

Check Model Files

$modelDir   = "<OCR_DIR>\models\GLM-OCR-GGUF"
$modelFile  = "$modelDir\GLM-OCR-Q8_0.gguf"
$mmprojFile = "$modelDir\mmproj-GLM-OCR-Q8_0.gguf"

$modelOk  = Test-Path $modelFile
$mmprojOk = Test-Path $mmprojFile

if ($modelOk -and $mmprojOk) {
    Write-Host "OK: GLM-OCR model files ready, skip Step 2"
    Write-Host "MODEL_STATUS=READY"
} else {
    if (-not $modelOk)  { Write-Host "ERROR: Missing GLM-OCR-Q8_0.gguf" }
    if (-not $mmprojOk) { Write-Host "ERROR: Missing mmproj-GLM-OCR-Q8_0.gguf" }
    Write-Host "MODEL_STATUS=MISSING"
    Write-Host "   Checked path: $modelDir"
}

Announce: ✅ Environment check complete. Execute steps as needed.

Step 1: Install / Update llama.cpp Vulkan

⬇️ Step 1: Downloading and installing llama.cpp Vulkan… (only when LLAMA_STATUS=MISSING/OUTDATED)

$tag      = "b8400"   # Replace with the latest tag from https://github.com/ggml-org/llama.cpp/releases/latest
$llamaDir = "<OCR_DIR>\llama.cpp"
$zip      = "$env:TEMP\llama-vulkan.zip"
$url      = "https://github.com/ggml-org/llama.cpp/releases/download/$tag/llama-$tag-bin-win-vulkan-x64.zip"

Write-Host "Downloading llama.cpp $tag ..."
Invoke-WebRequest -Uri $url -OutFile $zip

New-Item -ItemType Directory -Force -Path $llamaDir | Out-Null
Expand-Archive $zip -DestinationPath $llamaDir -Force
Remove-Item $zip
Write-Host "LLAMA_INSTALL=DONE"

Announce: ✅ llama.cpp installed. Continue to Step 2 to download models.

Step 2: Download GLM-OCR Models

📦 Step 2: Checking Python and downloading GLM-OCR models… (only when MODEL_STATUS=MISSING)

Note: Models are downloaded via Python's hf download (huggingface_hub) or modelscope. The script will auto-locate any existing Python installation; if none is found, Miniforge will be installed automatically to %USERPROFILE%\miniforge3 (no admin rights required).

First-time Download Notice (required reading when MODEL_STATUS=MISSING)

Announce the following to the user, then ask whether to proceed:

📥 First-time model download is approximately 1.5 GB
   (GLM-OCR-Q8_0.gguf ~950 MB + mmproj ~484 MB).
   Estimated download time:
   • 100 Mbps connection: ~2 minutes
   •  50 Mbps connection: ~4 minutes
   •  10 Mbps connection: ~20 minutes

   Downloads support resumption — if interrupted, re-running this step
   will automatically continue from where it left off.

   ✅ Ready — start automatic download
   📂 I prefer to download manually — skip automatic download

• User chooses automatic download → continue with Python check and download commands below
• User chooses manual download → jump to the "Manual Download Fallback" section at the end of this step

Check Disk Space

$drive = Split-Path "<OCR_DIR>" -Qualifier
$free  = (Get-PSDrive ($drive.TrimEnd(':'))).Free / 1GB
Write-Host "DISK_FREE=$([math]::Round($free,1))GB"
if ($free -lt 2) {
    Write-Host "DISK_STATUS=LOW"
    Write-Host "[WARN] Less than 2 GB available — download may fail"
} else {
    Write-Host "DISK_STATUS=OK"
}

Check Python

# ── Optional: if you know the Python path, fill it in; leave blank to auto-search ──
$customPythonExe = ""   # e.g. "C:\Python311\python.exe"
# ──────────────────────────────────────────────────────────────────────────────────

$pythonExe = $null

# 1. User-specified path
if ($customPythonExe -and (Test-Path $customPythonExe)) {
    $ver = & $customPythonExe --version 2>&1
    Write-Host "OK: Using specified Python: $customPythonExe -> $ver"
    $pythonExe = $customPythonExe
}

# 2. Search PATH
if (-not $pythonExe) {
    foreach ($cmd in @("python", "python3", "py")) {
        if (Get-Command $cmd -ErrorAction SilentlyContinue) {
            $ver = & $cmd --version 2>&1
            Write-Host "OK: Found Python in PATH: $cmd -> $ver"
            $pythonExe = (Get-Command $cmd).Source
            break
        }
    }
}

# 3. Scan common install directories
if (-not $pythonExe) {
    $searchPaths = @(
        "$env:USERPROFILE\miniforge3\python.exe",
        "$env:USERPROFILE\miniconda3\python.exe",
        "$env:USERPROFILE\anaconda3\python.exe",
        "$env:LOCALAPPDATA\Programs\Python\Python3*\python.exe",
        "C:\Python3*\python.exe"
    )
    foreach ($pattern in $searchPaths) {
        $found = Get-Item $pattern -ErrorAction SilentlyContinue | Select-Object -First 1
        if ($found) {
            $ver = & $found.FullName --version 2>&1
            Write-Host "OK: Found Python in common directory: $($found.FullName) -> $ver"
            $pythonExe = $found.FullName
            break
        }
    }
}

if ($pythonExe) {
    $env:PYTHON_EXE = $pythonExe
    Write-Host "PYTHON_OK"
} else {
    Write-Host "ERROR: Python not found. Install Miniforge or set `$customPythonExe"
    Write-Host "PYTHON_MISSING"
}

If Python is not found, install Miniforge:

$mf = "$env:TEMP\Miniforge3-Windows-x86_64.exe"
Invoke-WebRequest `
  -Uri "https://github.com/conda-forge/miniforge/releases/latest/download/Miniforge3-Windows-x86_64.exe" `
  -OutFile $mf
Start-Process $mf -ArgumentList "/S /D=$env:USERPROFILE\miniforge3" -Wait
Remove-Item $mf
$env:PYTHON_EXE = "$env:USERPROFILE\miniforge3\python.exe"
& $env:PYTHON_EXE --version
Write-Host "PYTHON_OK"

Download Models

Option A: hf download (recommended)

& $env:PYTHON_EXE -m pip install huggingface_hub -q

# For users in China: set mirror (skip if outside China)
$env:HF_ENDPOINT = "https://hf-mirror.com"

$modelDir = "<OCR_DIR>\models\GLM-OCR-GGUF"
New-Item -ItemType Directory -Force -Path $modelDir | Out-Null

hf download ggml-org/GLM-OCR-GGUF `
  --include "GLM-OCR-Q8_0.gguf" "mmproj-GLM-OCR-Q8_0.gguf" `
  --local-dir $modelDir

Write-Host "MODEL_DOWNLOAD=DONE"

Option B: ModelScope (alternative for users in China)

& $env:PYTHON_EXE -m pip install modelscope -q
& $env:PYTHON_EXE -c "
from modelscope.hub.file_download import model_file_download
import os
dest = r'<OCR_DIR>\models\GLM-OCR-GGUF'
os.makedirs(dest, exist_ok=True)
model_file_download('ggml-org/GLM-OCR-GGUF', file_path='GLM-OCR-Q8_0.gguf', local_dir=dest)
model_file_download('ggml-org/GLM-OCR-GGUF', file_path='mmproj-GLM-OCR-Q8_0.gguf', local_dir=dest)
print('MODEL_DOWNLOAD=DONE')
"

Verify:

$modelDir = "<OCR_DIR>\models\GLM-OCR-GGUF"
Get-Item "$modelDir\GLM-OCR-Q8_0.gguf", "$modelDir\mmproj-GLM-OCR-Q8_0.gguf" |
  Select-Object Name, @{N='MB';E={[math]::Round($_.Length/1MB,0)}}

Announce: ✅ Model download complete.

Manual Download Fallback

If automatic download repeatedly fails, guide the user to download manually and place files in the correct directory:

⚠️ Automatic download failed. Please manually download the following two files:

1. GLM-OCR-Q8_0.gguf (~950 MB)
   HuggingFace: https://huggingface.co/ggml-org/GLM-OCR-GGUF/resolve/main/GLM-OCR-Q8_0.gguf
   HF Mirror:   https://hf-mirror.com/ggml-org/GLM-OCR-GGUF/resolve/main/GLM-OCR-Q8_0.gguf
   ModelScope:  https://modelscope.cn/models/ggml-org/GLM-OCR-GGUF/resolve/master/GLM-OCR-Q8_0.gguf

2. mmproj-GLM-OCR-Q8_0.gguf (~484 MB)
   HuggingFace: https://huggingface.co/ggml-org/GLM-OCR-GGUF/resolve/main/mmproj-GLM-OCR-Q8_0.gguf
   HF Mirror:   https://hf-mirror.com/ggml-org/GLM-OCR-GGUF/resolve/main/mmproj-GLM-OCR-Q8_0.gguf
   ModelScope:  https://modelscope.cn/models/ggml-org/GLM-OCR-GGUF/resolve/master/mmproj-GLM-OCR-Q8_0.gguf

Once downloaded, place both files into:
   <OCR_DIR>\models\GLM-OCR-GGUF\

Then re-run the Verify command to confirm the files are intact before continuing to Step 3.

Step 3: Process Recognition Result

🔍 Step 3: Processing GLM-OCR recognition result…

Determine Input Source

Case A: User Provides a Local File Path

Extract the file path from the user's message, then call llama-cli directly:

# ── Fix encoding ──
chcp 65001 | Out-Null
[Console]::OutputEncoding = [System.Text.Encoding]::UTF8
$OutputEncoding = [System.Text.Encoding]::UTF8

$imgPath = "<file path extracted from user message>"
$m       = "<OCR_DIR>\models\GLM-OCR-GGUF\GLM-OCR-Q8_0.gguf"
$mm      = "<OCR_DIR>\models\GLM-OCR-GGUF\mmproj-GLM-OCR-Q8_0.gguf"

if (-not (Test-Path $imgPath)) {
    Write-Host "ERROR: File not found: $imgPath"
    exit 1
}

$cliExe = "<OCR_DIR>\llama.cpp\llama-cli.exe"
$result = & $cliExe `
  -m $m `
  --mmproj $mm `
  --image $imgPath `
  -p "Please recognize and extract all text from this image. Output the text content line by line, preserving the original layout." `
  -ngl 99 `
  --device Vulkan0 `
  -c 12000 `
  2>$null

Write-Host $result

Success criteria: stdout contains the recognized text content.

Case B: User Uploaded an Image via the Interface

OpenClaw saves uploaded images to a temporary path. Retrieve that path from context and call llama-cli the same way:

# ── Fix encoding ──
chcp 65001 | Out-Null
[Console]::OutputEncoding = [System.Text.Encoding]::UTF8
$OutputEncoding = [System.Text.Encoding]::UTF8

# imgPath is the temporary image path provided by OpenClaw in context
$imgPath = "<temporary image path provided by OpenClaw>"
$m       = "<OCR_DIR>\models\GLM-OCR-GGUF\GLM-OCR-Q8_0.gguf"
$mm      = "<OCR_DIR>\models\GLM-OCR-GGUF\mmproj-GLM-OCR-Q8_0.gguf"

if (-not (Test-Path $imgPath)) {
    Write-Host "ERROR: File not found: $imgPath"
    exit 1
}

$cliExe = "<OCR_DIR>\llama.cpp\llama-cli.exe"
$result = & $cliExe `
  -m $m `
  --mmproj $mm `
  --image $imgPath `
  -p "Please recognize and extract all text from this image. Output the text content line by line, preserving the original layout." `
  -ngl 99 `
  --device Vulkan0 `
  -c 12000 `
  2>$null

Write-Host $result

Success criteria: stdout contains the recognized text content.

Format Output

Once the recognized text is obtained, process it according to the user's intent:

Completion announcement:

✅ Recognition complete!
Let me know if you'd like to re-process, change the output format, or export to a file.

Troubleshooting

References

• llama.cpp Releases: https://github.com/ggml-org/llama.cpp/releases
• GLM-OCR GGUF: https://huggingface.co/ggml-org/GLM-OCR-GGUF