| | | 1 | | using Intentum.AI.Catalog; |
| | | 2 | | using Intentum.AI.Embeddings; |
| | | 3 | | using Intentum.Core.Behavior; |
| | | 4 | | using Intentum.Core.Intents; |
| | | 5 | | using Intentum.Core.Pipeline; |
| | | 6 | | |
| | | 7 | | namespace Intentum.AI.Models; |
| | | 8 | | |
| | | 9 | | /// <summary> |
| | | 10 | | /// AI inference step that classifies intent by comparing behavior embeddings against |
| | | 11 | | /// a catalog of known intent definitions using cosine similarity (nearest neighbor). |
| | | 12 | | /// Unlike LlmInferenceStep which always returns "AI-Inferred-Intent", this step |
| | | 13 | | /// returns the actual matched intent name from the catalog. |
| | | 14 | | /// </summary> |
| | | 15 | | public sealed class CatalogInferenceStep : IIntentInferenceStep |
| | | 16 | | { |
| | | 17 | | private readonly IIntentEmbeddingProvider _embeddingProvider; |
| | | 18 | | private readonly IntentCatalog _catalog; |
| | | 19 | | |
| | 2 | 20 | | public CatalogInferenceStep( |
| | 2 | 21 | | IIntentEmbeddingProvider embeddingProvider, |
| | 2 | 22 | | IntentCatalog catalog) |
| | 2 | 23 | | { |
| | 2 | 24 | | _embeddingProvider = embeddingProvider ?? throw new ArgumentNullException(nameof(embeddingProvider)); |
| | 2 | 25 | | _catalog = catalog ?? throw new ArgumentNullException(nameof(catalog)); |
| | 2 | 26 | | } |
| | | 27 | | |
| | | 28 | | /// <inheritdoc /> |
| | | 29 | | public IntentInferenceResult Infer(BehaviorSpace behaviorSpace, BehaviorVector vector) |
| | 2 | 30 | | { |
| | 2 | 31 | | var embeddings = vector.Dimensions.Keys |
| | 2 | 32 | | .Select(_embeddingProvider.Embed) |
| | 2 | 33 | | .ToList(); |
| | | 34 | | |
| | 2 | 35 | | return Classify(embeddings); |
| | 2 | 36 | | } |
| | | 37 | | |
| | | 38 | | /// <summary> |
| | | 39 | | /// Async version that properly awaits embedding calls. |
| | | 40 | | /// </summary> |
| | | 41 | | public async Task<IntentInferenceResult> InferAsync( |
| | | 42 | | BehaviorSpace behaviorSpace, |
| | | 43 | | BehaviorVector vector, |
| | | 44 | | CancellationToken cancellationToken = default) |
| | 0 | 45 | | { |
| | 0 | 46 | | var tasks = vector.Dimensions.Keys |
| | 0 | 47 | | .Select(k => _embeddingProvider.EmbedAsync(k, cancellationToken)); |
| | 0 | 48 | | var embeddings = (await Task.WhenAll(tasks)).ToList(); |
| | | 49 | | |
| | 0 | 50 | | return Classify(embeddings); |
| | 0 | 51 | | } |
| | | 52 | | |
| | | 53 | | private IntentInferenceResult Classify(List<IntentEmbedding> embeddings) |
| | 2 | 54 | | { |
| | 5 | 55 | | var vectorsWithData = embeddings.Where(e => e.Vector is { Count: > 0 }).ToList(); |
| | | 56 | | |
| | 2 | 57 | | string intentName = "Unknown"; |
| | 2 | 58 | | double bestScore = 0; |
| | 2 | 59 | | string? reasoning = null; |
| | | 60 | | |
| | 2 | 61 | | if (vectorsWithData.Count > 0) |
| | 2 | 62 | | { |
| | 5 | 63 | | var avgVector = AverageVectors(vectorsWithData.Select(e => e.Vector!).ToList()); |
| | 2 | 64 | | var match = _catalog.FindBestMatch(avgVector); |
| | | 65 | | |
| | 2 | 66 | | if (match.HasValue) |
| | 2 | 67 | | { |
| | 2 | 68 | | intentName = match.Value.Name; |
| | 2 | 69 | | bestScore = match.Value.Score; |
| | 2 | 70 | | reasoning = $"Nearest catalog match: {intentName} (similarity: {bestScore:F3})"; |
| | 2 | 71 | | } |
| | 2 | 72 | | } |
| | | 73 | | |
| | 2 | 74 | | if (Math.Abs(bestScore) < 1e-9 && embeddings.Count > 0) |
| | 1 | 75 | | { |
| | 2 | 76 | | bestScore = embeddings.Average(e => e.Score); |
| | 1 | 77 | | reasoning = "No catalog match found; using average embedding score"; |
| | 1 | 78 | | } |
| | | 79 | | |
| | 2 | 80 | | var signals = embeddings.Select(e => |
| | 3 | 81 | | new IntentSignal(Source: "catalog", Description: e.Source, Weight: e.Score)) |
| | 2 | 82 | | .ToList(); |
| | | 83 | | |
| | 2 | 84 | | return new IntentInferenceResult( |
| | 2 | 85 | | Name: intentName, |
| | 2 | 86 | | Score: bestScore, |
| | 2 | 87 | | Signals: signals, |
| | 2 | 88 | | Reasoning: reasoning |
| | 2 | 89 | | ); |
| | 2 | 90 | | } |
| | | 91 | | |
| | | 92 | | private static double[] AverageVectors(List<IReadOnlyList<double>> vectors) |
| | 2 | 93 | | { |
| | 2 | 94 | | var dim = vectors[0].Count; |
| | 2 | 95 | | var avg = new double[dim]; |
| | | 96 | | |
| | 12 | 97 | | foreach (var vec in vectors) |
| | 3 | 98 | | { |
| | 54 | 99 | | for (var i = 0; i < Math.Min(dim, vec.Count); i++) |
| | 24 | 100 | | avg[i] += vec[i]; |
| | 3 | 101 | | } |
| | | 102 | | |
| | 36 | 103 | | for (var i = 0; i < dim; i++) |
| | 16 | 104 | | avg[i] /= vectors.Count; |
| | | 105 | | |
| | 2 | 106 | | return avg; |
| | 2 | 107 | | } |
| | | 108 | | } |