Optimisation du RAG des ateliers : seuil de pertinence, recherche hybride, cache et overlap

- Seuil rag_min_score (defaut 0.30) : plus d'extraits hors-sujet injectes dans le prompt
- Recherche hybride : cosinus + bonus lexical (noms propres JdR mieux retrouves)
- Cache memoire du vector store (invalidation mtime) : plus de re-parse JSON par question
- Overlap de 80 tokens entre extraits RAG consecutifs (phrases a cheval retrouvables)
- Script de non-regression brain/scripts/sanity_rag_check.py

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

brain/scripts/sanity_rag_check.py

@@ -0,0 +1,84 @@
+# -*- coding: utf-8 -*-
+"""Sanity check temporaire : overlap du chunking + recherche hybride du vector store."""
+import sys
+from pathlib import Path
+
+sys.path.insert(0, str(Path(__file__).resolve().parents[1]))
+
+from app.application.chunking import chunk_text
+
+# --- 1. Chunking avec overlap ---
+paras = [f"Paragraphe {i} : " + ("lorem ipsum dolor sit amet " * 8) for i in range(12)]
+text = "\n\n".join(paras)
+
+no_overlap = chunk_text(text, target_tokens=200)
+with_overlap = chunk_text(text, target_tokens=200, overlap_tokens=40)
+
+assert len(with_overlap) >= len(no_overlap), "l'overlap ne doit pas réduire le nb de chunks"
+# Chaque chunk (sauf le 1er) doit commencer par la fin du précédent
+overlapped = 0
+for prev, cur in zip(with_overlap, with_overlap[1:]):
+    first_para = cur.split("\n\n")[0]
+    if first_para in prev:
+        overlapped += 1
+assert overlapped >= len(with_overlap) - 2, f"overlap absent: {overlapped}/{len(with_overlap)-1}"
+# Pas de chunk composé uniquement de l'overlap (dernier chunk dupliqué)
+assert with_overlap[-1] != with_overlap[-2], "dernier chunk = pur overlap (dupliqué)"
+# overlap_tokens=0 → comportement identique à l'ancien
+assert no_overlap == chunk_text(text, target_tokens=200, overlap_tokens=0)
+print(f"[OK] chunking : {len(no_overlap)} chunks sans overlap, {len(with_overlap)} avec ({overlapped} recouvrements)")
+
+# --- Paragraphe géant ---
+giant = "mot " * 2000
+sub = chunk_text(giant, target_tokens=300, overlap_tokens=50)
+assert len(sub) > 1
+print(f"[OK] paragraphe géant coupé en {len(sub)} sous-blocs")
+
+# --- 2. Vector store : hybride + seuil + cache ---
+import tempfile, os
+from app.infrastructure import vector_store
+
+with tempfile.TemporaryDirectory() as tmp:
+    vector_store._STORE_DIR = Path(tmp)
+    chunks = [
+        "Strahd von Zarovich règne sur la sombre vallée de Barovia depuis son château.",
+        "Les règles de combat utilisent un d20 plus le modificateur de caractéristique.",
+        "La taverne du village sert un ragoût de navets aux voyageurs fatigués.",
+    ]
+    # Vecteurs factices : chunk 0 et 1 proches de la query, chunk 2 orthogonal
+    vectors = [[1.0, 0.1, 0.0], [0.9, 0.4, 0.1], [0.0, 0.0, 1.0]]
+    vector_store.save("src1", chunks, vectors, pages=[10, 20, 30])
+
+    q = [1.0, 0.2, 0.0]
+    # Sans seuil ni texte : 3 résultats, ordre cosinus
+    r = vector_store.search(["src1"], q, top_k=10)
+    assert len(r) == 3 and r[0]["page"] == 10
+
+    # Avec seuil : le chunk orthogonal (cos~0) est écarté
+    r = vector_store.search(["src1"], q, top_k=10, min_score=0.30)
+    assert len(r) == 2, f"seuil non appliqué: {len(r)}"
+    print(f"[OK] seuil : 2/3 extraits gardés (orthogonal écarté)")
+
+    # Bonus lexical : la query mentionne « Strahd Barovia » → chunk 0 doit dominer
+    r = vector_store.search(["src1"], q, top_k=10, query_text="Parle-moi de Strahd et de Barovia", min_score=0.30)
+    assert r[0]["text"].startswith("Strahd"), r[0]["text"]
+    assert r[0]["score"] > vector_store._cosine(q, vectors[0]), "bonus lexical absent"
+    print(f"[OK] hybride : bonus lexical appliqué (score={r[0]['score']:.3f})")
+
+    # Le set "_words" mémoïsé ne doit PAS fuiter dans les résultats
+    assert all("_words" not in res for res in r)
+
+    # Cache : 2e recherche sert depuis la mémoire (même objet liste)
+    c1 = vector_store._load("src1")
+    c2 = vector_store._load("src1")
+    assert c1 is c2, "cache mtime inopérant"
+    # save() invalide le cache
+    vector_store.save("src1", chunks[:1], vectors[:1], pages=[10])
+    c3 = vector_store._load("src1")
+    assert len(c3) == 1, "cache non invalidé après save"
+    # delete() purge cache + fichier
+    vector_store.delete("src1")
+    assert vector_store._load("src1") == []
+    print("[OK] cache mémoire : hit, invalidation save, purge delete")
+
+print("\nTous les sanity checks passent.")