improve cache model

app.py

@@ -1,10 +1,10 @@
 import streamlit as st
-import faiss
 import numpy as np
 import json
-from sentence_transformers import SentenceTransformer
+from sentence_transformers import SentenceTransformer, util
 import time
 
+st.set_page_config(initial_sidebar_state="collapsed")
 # データを読み込む
 with open("data/qa_data.json", "r", encoding="utf-8") as f:
     data = json.load(f)
@@ -12,15 +12,27 @@ with open("data/qa_data.json", "r", encoding="utf-8") as f:
 questions = [item["question"] for item in data]
 answers = [item["answer"] for item in data]
 
-# 埋め込みモデルをロード
-model = SentenceTransformer("pkshatech/GLuCoSE-base-ja")
 
-# FAISSインデックスをロード
-index_q = faiss.read_index("data/faiss_question.index")
-index_a = faiss.read_index("data/faiss_answer.index")
+# Cache model ở level app
+@st.cache_resource
+def load_model():
+    return SentenceTransformer("pkshatech/GLuCoSE-base-ja")
+
+
+# Cache embeddings data
+@st.cache_data
+def load_embeddings():
+    return (
+        np.load("data/question_embeddings.npy"),
+        np.load("data/answer_embeddings.npy"),
+    )
+
+
+# Load model và embeddings một lần
+model = load_model()
+question_embeddings, answer_embeddings = load_embeddings()
 
 # サイドバー設定
-st.set_page_config(initial_sidebar_state="collapsed")
 with st.sidebar.expander("⚙️ 設定", expanded=False):
     threshold_q = st.slider("質問の類似度しきい値", 0.0, 1.0, 0.7, 0.01)
     threshold_a = st.slider("回答の類似度しきい値", 0.0, 1.0, 0.65, 0.01)
@@ -31,24 +43,37 @@ with st.sidebar.expander("⚙️ 設定", expanded=False):
 
 
 def search_answer(user_input):
-    """FAISSを使用して最適な回答を検索"""
-    user_embedding = model.encode([user_input]).astype(np.float32)
-
-    # 質問に対して検索
-    D_q, I_q = index_q.search(user_embedding, 1)
-    score_q = 1 / (1 + D_q[0][0])
+    """Tìm kiếm câu trả lời sử dụng cosine similarity"""
+    # Encode với batch_size và show_progress_bar=False để tăng tốc
+    user_embedding = model.encode(
+        [user_input],
+        convert_to_numpy=True,
+        batch_size=1,
+        show_progress_bar=False,
+        normalize_embeddings=True,  # Pre-normalize để tăng tốc cosine similarity
+    )
+
+    # Tính cosine similarity với câu hỏi
+    cos_scores_q = util.cos_sim(user_embedding, question_embeddings)[0]
+    best_q_idx = np.argmax(cos_scores_q)
+    score_q = cos_scores_q[best_q_idx]
 
     if score_q >= threshold_q:
-        # Replace \n with markdown line breaks
-        return answers[I_q[0][0]].replace("\n", "  \n"), f"質問にマッチ ({score_q:.2f})"
+        return (
+            answers[best_q_idx].replace("\n", "  \n"),
+            f"質問にマッチ ({score_q:.2f})",
+        )
 
-    # 回答に対して検索
-    D_a, I_a = index_a.search(user_embedding, 1)
-    score_a = 1 / (1 + D_a[0][0])
+    # Tính cosine similarity với câu trả lời
+    cos_scores_a = model.util.cos_sim(user_embedding, answer_embeddings)[0]
+    best_a_idx = np.argmax(cos_scores_a)
+    score_a = cos_scores_a[best_a_idx]
 
     if score_a >= threshold_a:
-        # Replace \n with markdown line breaks
-        return answers[I_a[0][0]].replace("\n", "  \n"), f"回答にマッチ ({score_a:.2f})"
+        return (
+            answers[best_a_idx].replace("\n", "  \n"),
+            f"回答にマッチ ({score_a:.2f})",
+        )
 
     return "申し訳ありませんが、ご質問の答えを見つけることができませんでした。もう少し詳しく説明していただけますか？", "一致なし"
 
@@ -81,6 +106,7 @@ if user_input := st.chat_input("質問を入力してください:"):
 
     with st.spinner("考え中... お待ちください。"):
         answer, info = search_answer(user_input)
+        print(info)
 
     with st.chat_message("assistant"):
         response_placeholder = st.empty()

data/answer_embeddings.npy

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:925632dc69ab4df0223970df60cc9054dd46e2958e597e6998514bd3b33fc703
+size 67712

data/question_embeddings.npy

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:45127b6e8615f93324b2debb37305b93d3963c5f91f054f9c56def8cd00c1ca5
+size 67712

preprocess.py

@@ -1,5 +1,4 @@
 import json
-import faiss
 import numpy as np
 from sentence_transformers import SentenceTransformer
 
@@ -14,20 +13,13 @@ answers = [item["answer"] for item in data]
 model = SentenceTransformer("pkshatech/GLuCoSE-base-ja")
 
 # Tạo embedding cho câu hỏi và câu trả lời
-question_embeddings = model.encode(questions)
-answer_embeddings = model.encode(answers)
+question_embeddings = model.encode(questions, convert_to_numpy=True)
+answer_embeddings = model.encode(answers, convert_to_numpy=True)
 
-# Lưu FAISS index
-dim = question_embeddings.shape[1]
-index_q = faiss.IndexFlatL2(dim)
-index_a = faiss.IndexFlatL2(dim)
-
-index_q.add(np.array(question_embeddings).astype(np.float32))
-index_a.add(np.array(answer_embeddings).astype(np.float32))
-
-faiss.write_index(index_q, "faiss_question.index")
-faiss.write_index(index_a, "faiss_answer.index")
+# Lưu embedding dưới dạng numpy array
+np.save("data/question_embeddings.npy", question_embeddings)
+np.save("data/answer_embeddings.npy", answer_embeddings)
 
 # Lưu dữ liệu gốc
-with open("qa_data.json", "w", encoding="utf-8") as f:
+with open("data/qa_data.json", "w", encoding="utf-8") as f:
     json.dump(data, f, ensure_ascii=False, indent=2)