Upload 3 files

.gitattributes

@@ -34,3 +34,4 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 *.zst filter=lfs diff=lfs merge=lfs -text
 *tfevents* filter=lfs diff=lfs merge=lfs -text
 vectorstore/ruri-large/index.faiss filter=lfs diff=lfs merge=lfs -text
+vectorstore/static-embedding-japanese/index.faiss filter=lfs diff=lfs merge=lfs -text

app.py

@@ -0,0 +1,181 @@
+import os
+import gradio as gr
+from huggingface_hub import hf_hub_download
+from langchain.chains import RetrievalQA
+from langchain_community.vectorstores import FAISS
+from langchain_community.embeddings import HuggingFaceEmbeddings
+from langchain_community.llms import LlamaCpp
+
+
+REPO_ID = "WariHima/sarashina2.2-1b-instruct-v0.1-Q4_K_M-GGUF"
+FILENAME = "sarashina2.2-1b-instruct-v0.1-q4_k_m.gguf"
+
+
+def get_model_path():
+    return hf_hub_download(
+        repo_id=REPO_ID,
+        filename=FILENAME,
+        repo_type="model",
+    )
+
+
+GGUF_MODEL_PATH = get_model_path()
+VECTOR_DB_PATH = "./vectorstore/static-embedding-japanese"
+EMBEDDING_MODEL = "hotchpotch/static-embedding-japanese"
+
+
+class RAGSystem:
+    def __init__(self):
+        self.vectorstore = None
+        self.qa_chain = None
+        self.setup_models()
+
+    def setup_models(self):
+        self.embeddings = HuggingFaceEmbeddings(
+            model_name=EMBEDDING_MODEL,
+            model_kwargs={"device": "cpu"},
+        )
+
+        try:
+            self.load_vectorstore()
+        except Exception as e:
+            print(f"ベクトルDBの読み込みに失敗しました: {str(e)}")
+
+        try:
+            self.llm = LlamaCpp(
+                model_path=GGUF_MODEL_PATH,
+                temperature=0.7,
+                max_tokens=512,
+                n_ctx=2048,  # コンテキスト長
+                n_threads=8,  # 使用するCPUスレッド数
+                n_gpu_layers=-1,  # 可能であればGPUレイヤーを全て使用
+                verbose=False,
+                streaming=True,
+                model_kwargs={"f16_kv": True},
+            )
+
+            if self.vectorstore:
+                self.setup_qa_chain()
+        except Exception as e:
+            print(f"LLMの読み込みに失敗しました: {str(e)}")
+
+    def load_vectorstore(self):
+        if os.path.exists(VECTOR_DB_PATH):
+            self.vectorstore = FAISS.load_local(
+                VECTOR_DB_PATH,
+                self.embeddings,
+                allow_dangerous_deserialization=True,
+            )
+            if self.llm:
+                self.setup_qa_chain()
+            return True
+        return False
+
+    def setup_qa_chain(self):
+        if self.vectorstore and self.llm:
+            self.qa_chain = RetrievalQA.from_chain_type(
+                llm=self.llm,
+                chain_type="stuff",
+                retriever=self.vectorstore.as_retriever(search_kwargs={"k": 3}),
+            )
+            return True
+        return False
+
+    def answer_question_stream(self, question):
+        if not self.qa_chain:
+            if not self.vectorstore:
+                yield "ベクトルDBが読み込まれていません。"
+                return
+            if not self.llm:
+                yield "LLMモデルが読み込まれていません。"
+                return
+            yield "QAチェーンの初期化に失敗しました。"
+            return
+
+        try:
+            docs = self.vectorstore.similarity_search(question, k=3)
+            context = "\n\n".join([doc.page_content for doc in docs])
+
+            prompt = f"""与えられた文書を用いて、質問に対する適切な応答を書きなさい。
+文書: {context}
+質問: {question}
+応答: """
+
+            response = ""
+            for chunk in self.llm._stream(prompt):
+                if isinstance(chunk, str):
+                    response += chunk
+                else:
+                    response += chunk.text
+                yield response
+
+        except Exception as e:
+            yield f"回答生成中にエラーが発生しました: {str(e)}"
+
+    def get_system_status(self):
+        status = list()
+        if os.path.exists(GGUF_MODEL_PATH):
+            model_size = os.path.getsize(GGUF_MODEL_PATH) / (1024 * 1024 * 1024)
+            status.append(
+                f"✅ LLMモデル: {os.path.basename(GGUF_MODEL_PATH)} ({model_size:.2f} GB)"
+            )
+        else:
+            status.append(f"❌ LLMモデル: {GGUF_MODEL_PATH} が見つかりません")
+
+        if os.path.exists(VECTOR_DB_PATH):
+            status.append(f"✅ ベクトルDB: {VECTOR_DB_PATH}")
+        else:
+            status.append(f"❌ ベクトルDB: {VECTOR_DB_PATH} が見つかりません")
+
+        status.append(f"✅ 埋め込みモデル: {EMBEDDING_MODEL}")
+
+        if self.qa_chain:
+            status.append("✅ RAGシステム: 準備完了")
+        else:
+            status.append("❌ RAGシステム: 初期化されていません")
+
+        return "\n".join(status)
+
+
+rag_system = RAGSystem()
+
+with gr.Blocks(title="RAGデモアプリ") as demo:
+    gr.Markdown("# 🎇 Sake RAG デモアプリ")
+    gr.Markdown(
+        "醸造協会誌5年分のデータをベクトルDBとして保持した1B級の小型言語モデルです"
+    )
+
+    with gr.Row():
+        with gr.Column(scale=1):
+            refresh_button = gr.Button("システム状態を更新", variant="secondary")
+            status_output = gr.Textbox(
+                label="システム状態",
+                value=rag_system.get_system_status(),
+                interactive=False,
+                lines=5,
+            )
+
+        with gr.Column(scale=2):
+            question_input = gr.Textbox(
+                label="質問を入力してください",
+                placeholder="質問を入力してください",
+                lines=2,
+            )
+            submit_button = gr.Button("質問する", variant="primary")
+            answer_output = gr.Textbox(label="回答", interactive=False, lines=10)
+
+    refresh_button.click(
+        fn=rag_system.get_system_status,
+        inputs=[],
+        outputs=[status_output],
+    )
+
+    submit_button.click(
+        fn=rag_system.answer_question_stream,
+        inputs=[question_input],
+        outputs=[answer_output],
+    )
+
+
+if __name__ == "__main__":
+    demo.launch()

vectorstore/static-embedding-japanese/index.faiss

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:598f40e8715d934190631a8298a58371a957fddba4ba4572b83f6ec9676e85af
+size 7405613

vectorstore/static-embedding-japanese/index.pkl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ae8fff1ec1e0adda6565cea84064a604ae345e751c1dc522aca9185b23bc53d7
+size 826595