Text & image generation

// Ask a vision-language model about an image, then ask a follow-up question
// in the same conversation. Answers arrive as an AsyncThrowingStream.

import Foundation
import SlabSwift

func askAboutDocument(modelPath: String, imagePath: String, license: String) async throws {
    let config = try SlabConfig(license: license)
    let session = try SlabSession(config: config, path: modelPath, contextLength: 2048)

    let input = SlabInput()
    input.maxTokens = 64
    input.temperature = 0
    input.setImageTiled(true, longestEdge: 1536)   // large images are split into tiles, not shrunk

    let output = SlabOutput()
    input.addMessage(role: "user", text: "What is this document?", image: try SlabFrame(contentsOf: imagePath))
    for try await piece in session.generate(input, into: output) {
        print(piece, terminator: "")
    }
    print()

    // Send the conversation again with the answer and a new question.
    // The shared beginning is reused, not recomputed.
    input.addMessage(role: "assistant", text: output.fullText)
    input.addMessage(role: "user", text: "What is the total amount?")
    for try await piece in session.generate(input, into: output) {
        print(piece, terminator: "")
    }
    print()
    print("prompt \(output.promptTokens) tokens (\(output.cachedPromptTokens) reused), generated \(output.generatedTokens)")

    try session.reset()   // forget the conversation before an unrelated one
}