//
//  SemanticIndexer.swift
//  Focale
//
//  Author: Simon-Pierre Boucher
//  Contact: contact@spboucher.ai
//
//  Stage 2 — Foundation Models, on candidates only (CLAUDE.md §5).
//  The math that rules everything: 40 000 photos × 1–3 s = 11–33 h, serial.
//  So this actor runs one inference at a time, only on photos that earn it.
//
//  Note: direct image input in the prompt arrives with the iOS 27 SDK;
//  until then stage 2 reasons over stage-1 signals + capture context,
//  which is already rich for Focale-taken photos.
//

import Foundation
import FoundationModels

/// Why a photo earned a stage-2 pass. Anything else stays Vision-only.
enum SemanticCandidateReason: Sendable {
    case capturedInFocale     // context already present: minimal cost, maximal value
    case userRequested        // the user opened the photo and asked
    case searchDeepening      // stage-1 search found nothing; deepen a small subset
    case markedImportant      // favorite, shared, frequently opened
}

/// Inputs handed to the model — stage-1 signals, never the photo library.
struct SemanticCandidate: Sendable {
    var localIdentifier: String
    var ocrExcerpt: String?
    var classificationLabels: [String]
    var contextJSON: String?
}

actor SemanticIndexer {

    static var isModelAvailable: Bool {
        SystemLanguageModel.default.availability == .available
    }

    /// Serialized on the Neural Engine anyway (CLAUDE.md §2) — one session,
    /// one inference at a time.
    private var session: LanguageModelSession?

    /// Returns nil on guardrail refusal — a normal state, no error surfaced
    /// (CLAUDE.md §8). The photo stays indexed by stage 1.
    func analyze(_ candidate: SemanticCandidate) async -> PhotoSemantics? {
        guard Self.isModelAvailable else { return nil }

        let session = self.session ?? LanguageModelSession(instructions: """
            Tu analyses les signaux extraits d'une photo pour la rendre cherchable.
            Réponds uniquement à partir des signaux fournis. Sur incertitude, \
            reste vague plutôt que d'inventer. Ne devine jamais un chiffre, \
            un montant ou une date : ils viennent de l'OCR, pas de toi.
            """)
        self.session = session

        var prompt = "Signaux de la photo :\n"
        if let ocr = candidate.ocrExcerpt, !ocr.isEmpty {
            prompt += "Texte OCR (données réelles) :\n\(String(ocr.prefix(1200)))\n"
        }
        if !candidate.classificationLabels.isEmpty {
            prompt += "Classification : \(candidate.classificationLabels.joined(separator: ", "))\n"
        }
        if let context = candidate.contextJSON {
            prompt += "Contexte de capture : \(context)\n"
        }

        do {
            let response = try await session.respond(
                to: prompt,
                generating: PhotoSemantics.self
            )
            return response.content
        } catch {
            // Guardrail refusal or transient failure: treat as normal,
            // reset the session so one bad turn doesn't poison the next.
            self.session = nil
            return nil
        }
    }
}