Skip to content

Code graph and memory for your coding agents, in one Rust program. For Claude Code, Codex, Cursor and any MCP client.

3.1 KB
Raw Blame
import Foundation
import AppKit
import Vision
import PDFKit
import ImageIO

enum DocumentError: Error { case invalid, tooLarge }
let limit = 512 * 1024
func recognize(_ image: CGImage) throws -> String {
    let request = VNRecognizeTextRequest()
    request.recognitionLevel = .accurate
    request.usesLanguageCorrection = true
    try VNImageRequestHandler(cgImage: image).perform([request])
    return (request.results ?? []).compactMap { $0.topCandidates(1).first?.string }.joined(separator: "\n")
}
do {
    guard CommandLine.arguments.count == 2 else { throw DocumentError.invalid }
    let url = URL(fileURLWithPath: CommandLine.arguments[1])
    let values = try url.resourceValues(forKeys: [.fileSizeKey, .isRegularFileKey])
    guard values.isRegularFile == true, let size = values.fileSize, size <= 16 * 1024 * 1024 else { throw DocumentError.tooLarge }
    let ext = url.pathExtension.lowercased()
    var text = ""
    var usedOCR = false
    if ext == "pdf" {
        guard let document = PDFDocument(url: url), document.pageCount <= 100 else { throw DocumentError.invalid }
        for index in 0..<document.pageCount {
            guard let page = document.page(at: index) else { throw DocumentError.invalid }
            let recognized: String = try autoreleasepool {
                if let original = page.string, !original.trimmingCharacters(in: .whitespacesAndNewlines).isEmpty { return original }
                let bounds = page.bounds(for: .mediaBox)
                let scale = min(2.0, 2400.0 / max(bounds.width, bounds.height))
                let image = page.thumbnail(of: NSSize(width: bounds.width * scale, height: bounds.height * scale), for: .mediaBox)
                guard let cg = image.cgImage(forProposedRect: nil, context: nil, hints: nil) else { throw DocumentError.invalid }
                usedOCR = true
                return try recognize(cg)
            }
            text += "\n[Page \(index + 1)]\n" + recognized
            if text.utf8.count > limit { throw DocumentError.tooLarge }
        }
    } else if ext == "doc" || ext == "rtf" {
        let content = try Data(contentsOf: url)
        let kind: NSAttributedString.DocumentType = ext == "rtf" ? .rtf : .docFormat
        text = try NSAttributedString(data: content, options: [.documentType: kind], documentAttributes: nil).string
    } else {
        guard let source = CGImageSourceCreateWithURL(url as CFURL, nil),
              let image = CGImageSourceCreateThumbnailAtIndex(source, 0, [kCGImageSourceCreateThumbnailFromImageAlways: true, kCGImageSourceThumbnailMaxPixelSize: 4096, kCGImageSourceCreateThumbnailWithTransform: true] as CFDictionary) else { throw DocumentError.invalid }
        text = try recognize(image)
        usedOCR = true
    }
    guard text.utf8.count <= limit, !text.trimmingCharacters(in: .whitespacesAndNewlines).isEmpty else { throw DocumentError.invalid }
    let result: [String: Any] = ["text": text, "parser": "macOS Vision/PDFKit", "ocr": usedOCR]
    FileHandle.standardOutput.write(try JSONSerialization.data(withJSONObject: result))
} catch {
    FileHandle.standardError.write(Data("Document conversion failed\n".utf8))
    exit(1)
}