manager-service/ManagerService/Services/GoogleEmbeddingService.cs

using ManagerService.Data;
using Microsoft.Extensions.Configuration;
using System;
using System.Collections.Generic;
using System.Linq;
using System.Net.Http;
using System.Net.Http.Json;
using System.Text.Json.Serialization;
using System.Threading;
using System.Threading.Tasks;

namespace ManagerService.Services
{
    /// <summary>
    /// Embeddings via l'endpoint OpenAI-compatible de Google — même hôte et même clé
    /// que le chat (voir Startup), donc aucune configuration supplémentaire.
    /// </summary>
    public class GoogleEmbeddingService : IEmbeddingService
    {
        private const string Model = "gemini-embedding-001";
        private const string Endpoint = "https://generativelanguage.googleapis.com/v1beta/openai/embeddings";

        /// <summary>
        /// L'API accepte des lots plus gros, mais un lot trop large fait échouer les ~150 morceaux
        /// d'un document d'un coup sur un seul dépassement de limite de jetons.
        /// </summary>
        private const int BatchSize = 50;

        private readonly IHttpClientFactory _httpClientFactory;
        private readonly string _apiKey;

        public int Dimensions => ContentEmbedding.Dimensions;

        public GoogleEmbeddingService(IHttpClientFactory httpClientFactory, IConfiguration configuration)
        {
            _httpClientFactory = httpClientFactory;
            _apiKey = configuration["AI:ApiKey"];
        }

        public async Task<float[]> EmbedAsync(string text, CancellationToken cancellationToken = default)
        {
            var result = await EmbedBatchAsync(new[] { text }, cancellationToken);
            return result[0];
        }

        public async Task<IReadOnlyList<float[]>> EmbedBatchAsync(IReadOnlyList<string> texts, CancellationToken cancellationToken = default)
        {
            if (texts == null || texts.Count == 0)
                return Array.Empty<float[]>();

            var results = new List<float[]>(texts.Count);

            for (var offset = 0; offset < texts.Count; offset += BatchSize)
            {
                var batch = texts.Skip(offset).Take(BatchSize).ToArray();
                results.AddRange(await EmbedOneBatchAsync(batch, cancellationToken));
            }

            return results;
        }

        private async Task<List<float[]>> EmbedOneBatchAsync(string[] batch, CancellationToken cancellationToken)
        {
            var client = _httpClientFactory.CreateClient();
            client.DefaultRequestHeaders.Authorization =
                new System.Net.Http.Headers.AuthenticationHeaderValue("Bearer", _apiKey);

            // dimensions est obligatoire : le modèle renvoie 3072 valeurs par défaut,
            // ce que la colonne vector(768) refuserait.
            var payload = new EmbeddingRequest
            {
                Model = Model,
                Input = batch,
                Dimensions = Dimensions
            };

            var response = await client.PostAsJsonAsync(Endpoint, payload, cancellationToken);

            if (!response.IsSuccessStatusCode)
            {
                var body = await response.Content.ReadAsStringAsync(cancellationToken);
                throw new InvalidOperationException(
                    $"Embedding API error ({(int)response.StatusCode}): {body}");
            }

            var parsed = await response.Content.ReadFromJsonAsync<EmbeddingResponse>(cancellationToken: cancellationToken);

            if (parsed?.Data == null || parsed.Data.Count != batch.Length)
                throw new InvalidOperationException(
                    $"Embedding API returned {parsed?.Data?.Count ?? 0} vectors for {batch.Length} texts.");

            // Vérifié le 2026-08-09 : Google ne renvoie pas le champ "index" du contrat
            // OpenAI — chaque item ne porte que "embedding" et "object". Trier dessus
            // donnerait un tri sur des zéros. L'ordre du tableau suit celui des entrées.
            return parsed.Data
                .Select(d => Normalize(d.Embedding))
                .ToList();
        }

        /// <summary>
        /// Normalise le vecteur en norme L2.
        /// gemini-embedding-001 ne renvoie des vecteurs normalisés que sur sa dimension
        /// native (3072). Toute dimension réduite sort brute, et la distance cosinus
        /// de pgvector suppose des vecteurs normalisés — sans ça, le classement des
        /// résultats est faussé sans qu'aucune erreur ne soit levée.
        /// </summary>
        private float[] Normalize(float[] vector)
        {
            if (vector == null || vector.Length != Dimensions)
                throw new InvalidOperationException(
                    $"Embedding API returned {vector?.Length ?? 0} dimensions, expected {Dimensions}.");

            double sumOfSquares = 0;
            foreach (var value in vector)
                sumOfSquares += (double)value * value;

            var norm = Math.Sqrt(sumOfSquares);
            if (norm == 0)
                return vector;

            var normalized = new float[vector.Length];
            for (var i = 0; i < vector.Length; i++)
                normalized[i] = (float)(vector[i] / norm);

            return normalized;
        }

        private class EmbeddingRequest
        {
            [JsonPropertyName("model")]
            public string Model { get; set; }

            [JsonPropertyName("input")]
            public string[] Input { get; set; }

            [JsonPropertyName("dimensions")]
            public int Dimensions { get; set; }
        }

        private class EmbeddingResponse
        {
            [JsonPropertyName("data")]
            public List<EmbeddingData> Data { get; set; }
        }

        private class EmbeddingData
        {
            [JsonPropertyName("embedding")]
            public float[] Embedding { get; set; }
        }
    }
}