Semantic caching

import numpy as np
 
# Pretend embeddings (an embedding model makes these in real life).
EMB = {
    "How do I reset my password?":          [0.90, 0.10, 0.00],
    "how can i reset my password":          [0.88, 0.12, 0.02],
    "I forgot my password, how to reset?":  [0.85, 0.15, 0.05],
    "How do I reset my router?":            [0.70, 0.10, 0.60],
    "What are your opening hours?":         [0.00, 0.95, 0.10],
}
def cos(a, b):
    a, b = np.array(a), np.array(b)
    return float(a @ b / (np.linalg.norm(a) * np.linalg.norm(b)))
 
class SemanticCache:
    def __init__(self, threshold):
        self.threshold, self.entries = threshold, []          # (embedding, answer)
    def get(self, q):
        best = max(self.entries, key=lambda e: cos(EMB[q], e[0]), default=None)
        if best and cos(EMB[q], best[0]) >= self.threshold:
            return best[1]
        return None
    def put(self, q, answer):
        self.entries.append((EMB[q], answer))
 
cache = SemanticCache(threshold=0.95)
cache.put("How do I reset my password?", "Go to Settings > Security > Reset password.")
for q in ["how can i reset my password", "I forgot my password, how to reset?",
          "How do I reset my router?", "What are your opening hours?"]:
    hit = cache.get(q)
    print(f"{q:38} -> {'HIT: ' + hit if hit else 'miss (call the model)'}")
print("router similarity:", round(cos(EMB['How do I reset my router?'], EMB['How do I reset my password?']), 2),
      "<- with a 0.75 threshold this would wrongly get the password answer")

how can i reset my password            -> HIT: Go to Settings > Security > Reset password.
I forgot my password, how to reset?    -> HIT: Go to Settings > Security > Reset password.
How do I reset my router?              -> miss (call the model)
What are your opening hours?           -> miss (call the model)
router similarity: 0.76 <- with a 0.75 threshold this would wrongly get the password answer