wahnsinn vibe

backend/apps/ai_shop/__init__.py

@@ -0,0 +1,111 @@
+import json
+import re
+
+from fastapi import APIRouter, Depends
+from pydantic import BaseModel
+from sqlalchemy import text
+from sqlalchemy.orm import Session
+
+from core.db import get_db
+from core.i18n import pick
+from core.redis_client import redis_client
+
+from apps.ai_core.ollama_client import get_llm
+
+router = APIRouter()
+
+
+class SearchIn(BaseModel):
+    query: str
+    limit: int = 12
+
+
+_STOP = {
+    "ich", "suche", "brauche", "will", "möchte", "eine", "einen", "ein",
+    "die", "der", "das", "und", "oder", "mit", "für", "zum", "zur",
+    "i", "am", "looking", "for", "want", "need", "a", "an", "the",
+    "of", "to", "with", "etwas", "some", "nach",
+}
+
+_SYN = {
+    "pulli": ["pullover", "sweater"],
+    "shirt": ["t-shirt", "tshirt"],
+    "hose": ["pants", "jeans"],
+    "warme": ["warm"],
+    "warm": ["warme"],
+    "grüner": ["grün", "green"],
+    "grüne": ["grün", "green"],
+    "grünes": ["grün", "green"],
+    "blauer": ["blau", "blue"],
+    "blaue": ["blau", "blue"],
+    "blaues": ["blau", "blue"],
+    "wandern": ["wander", "hiking"],
+}
+
+
+def _tokenize(s: str) -> list[str]:
+    tokens = [t.lower() for t in re.findall(r"[\wäöüß]+", s, flags=re.UNICODE)]
+    expanded: list[str] = []
+    for t in tokens:
+        if t in _STOP or len(t) < 2:
+            continue
+        expanded.append(t)
+        expanded.extend(_SYN.get(t, []))
+    return expanded
+
+
+def _keyword_score(product: dict, tokens: list[str]) -> float:
+    if not tokens:
+        return 0.0
+    haystack = " ".join([
+        pick(product.get("name", {}), "de").lower(),
+        pick(product.get("name", {}), "en").lower(),
+        pick(product.get("description", {}), "de").lower(),
+        pick(product.get("description", {}), "en").lower(),
+        " ".join(str(v).lower() for v in (product.get("attributes") or {}).values()),
+        product.get("sku", "").lower(),
+    ])
+    hits = sum(1 for t in tokens if t in haystack)
+    return hits / len(tokens)
+
+
+@router.post("/search")
+def ki_search(body: SearchIn, db: Session = Depends(get_db)):
+    """Hybrid product search: embedding similarity + keyword boost."""
+    q = body.query.strip()
+    if not q:
+        return {"query": q, "products": []}
+
+    emb = get_llm().embed(q)
+    # Pull a larger candidate pool, then re-rank with keyword boost
+    pool_size = max(body.limit * 3, 20)
+    rows = db.execute(
+        text(
+            """
+            SELECT source_id, 1 - (embedding <=> (:q)::vector) AS score
+            FROM ai_documents
+            WHERE source_type = 'product'
+            ORDER BY embedding <=> (:q)::vector
+            LIMIT :lim
+            """
+        ),
+        {"q": emb, "lim": pool_size},
+    ).mappings().all()
+
+    tokens = _tokenize(q)
+    candidates: list[dict] = []
+    for r in rows:
+        raw = redis_client.get(f"product:{r['source_id']}")
+        if not raw:
+            continue
+        d = json.loads(raw)
+        emb_s = float(r["score"])
+        kw_s = _keyword_score(d, tokens)
+        # Combined score: 60% embedding, 40% keyword (but keyword zeroing-out boosts ordering)
+        d["_score"] = round(0.6 * emb_s + 0.4 * kw_s, 4)
+        d["_emb"] = round(emb_s, 4)
+        d["_kw"] = round(kw_s, 4)
+        candidates.append(d)
+
+    candidates.sort(key=lambda p: p["_score"], reverse=True)
+    return {"query": q, "products": candidates[: body.limit]}