//
//  SemanticIndex.swift
//  Poche
//
//  Author:  Simon-Pierre Boucher
//  Contact: contact@spboucher.ai
//

import Foundation
import NaturalLanguage

/// Local semantic ranking over the user's data. NLEmbedding runs entirely
/// on-device — no service is involved, not even for indexing (CLAUDE.md §7).
///
/// Constraint: distances are computed per query over the whole corpus.
/// Fine at personal-notes scale; precompute stored vectors before the
/// corpus grows past a few thousand documents.
@MainActor
final class SemanticIndex {
    private let embedding = NLEmbedding.sentenceEmbedding(for: .french)

    func rank(query: String, in documents: [SearchDocument], limit: Int = 3) -> [SearchDocument] {
        guard !documents.isEmpty else { return [] }

        if let embedding {
            let scored = documents.map { document in
                (document, embedding.distance(between: query, and: document.text, distanceType: .cosine))
            }
            return scored
                .sorted { $0.1 < $1.1 }
                .prefix(limit)
                .map(\.0)
        }

        // Keyword fallback when the French sentence embedding asset is
        // not present on the device.
        let needles = query.lowercased().split(separator: " ").map(String.init)
        let scored = documents.map { document in
            (document, needles.count(where: { document.text.lowercased().contains($0) }))
        }
        return scored
            .filter { $0.1 > 0 }
            .sorted { $0.1 > $1.1 }
            .prefix(limit)
            .map(\.0)
    }
}