feat: vectorsearch-based QA

app.py

@@ -1,4 +1,114 @@
+from embeddings import KorRobertaEmbeddings
+
 import streamlit as st
+from streamlit import session_state as sst
+
+from langchain_core.runnables import (
+    RunnablePassthrough,
+    RunnableParallel,
+)
+
+PINECONE_API_KEY = st.secrets["PINECONE_API_KEY"]
+
+
+def create_or_get_pinecone_index(index_name: str, dimension: int = 768):
+    from pinecone import Pinecone, ServerlessSpec
+
+    client = Pinecone(api_key=PINECONE_API_KEY)
+    if index_name in [index["name"] for index in client.list_indexes()]:
+        pc_index = client.Index(index_name)
+        print("☑️ Got the existing Pinecone index")
+    else:
+        client.create_index(
+            name=index_name,
+            dimension=dimension,
+            metric="cosine",
+            spec=ServerlessSpec("aws", "us-west-2"),
+        )
+        pc_index = client.Index(index_name)
+        print("☑️ Created a new Pinecone index")
+
+    print(pc_index.describe_index_stats())
+    return pc_index
+
+
+def get_pinecone_vectorstore(
+    index_name: str,
+    embedding_fn=KorRobertaEmbeddings(),
+    dimension: int = 768,
+    namespace: str = None,
+):
+    from langchain_pinecone import Pinecone
+
+    index = create_or_get_pinecone_index(
+        index_name,
+        dimension,
+    )
+    vs = Pinecone(
+        index,
+        embedding_fn,
+        pinecone_api_key=PINECONE_API_KEY,
+        index_name=index_name,
+        namespace=namespace,
+    )
+    print(vs)
+    return vs
+
+
+def build_pinecone_retrieval_chain(vectorstore):
+    retriever = vectorstore.as_retriever()
+    rag_chain_with_source = RunnableParallel(
+        {"context": retriever, "question": RunnablePassthrough()}
+    )
+
+    return rag_chain_with_source
+
+
+@st.cache_resource
+def get_pinecone_retrieval_chain(collection_name):
+    print("☑️ Building a new pinecone retrieval chain...")
+    embed_fn = KorRobertaEmbeddings()
+    pinecone_vectorstore = get_pinecone_vectorstore(
+        index_name=collection_name,
+        embedding_fn=embed_fn,
+        dimension=768,
+        namespace="0221",
+    )
+
+    chain = build_pinecone_retrieval_chain(pinecone_vectorstore)
+    return chain
+
+
+def rerun():
+    st.rerun()
+
+
+st.title("이노션 데모")
+
+with st.spinner("환경 설정 중"):
+    sst.retrieval_chain = get_pinecone_retrieval_chain(
+        collection_name="innocean",
+    )
+
+if prompt := st.chat_input("정보 검색"):
+
+    # Display user message in chat message container
+    with st.chat_message("human"):
+        st.markdown(prompt)
+
+    # Get assistant response
+    outputs = sst.retrieval_chain.invoke(prompt)
+    print(outputs)
+    retrieval_docs = outputs["context"]
+
+    # Display assistant response in chat message container
+    with st.chat_message("assistant"):
+        st.markdown(retrieval_docs[0].metadata["answer"])
 
-x = st.slider('Select a value')
-st.write(x, 'squared is', x * x)
+        with st.expander("출처 보기", expanded=True):
+            st.info(f"출처 페이지: {retrieval_docs[0].metadata['page']}")
+            st.markdown(retrieval_docs[0].metadata["source_passage"])
+            # tabs = st.tabs([f"doc{i}" for i in range(len(retrieval_docs))])
+            # for i in range(len(retrieval_docs)):
+            #     tabs[i].write(retrieval_docs[i].page_content)
+            #     tabs[i].write(retrieval_docs[i].metadata)

embeddings.py

@@ -0,0 +1,37 @@
+from typing import List
+
+import os
+from langchain_core.embeddings import Embeddings
+from transformers import AutoModel, AutoTokenizer
+
+os.environ["TOKENIZERS_PARALLELISM"] = "false"
+
+
+def get_roberta_embeddings(sentences: List[str]):
+    """
+    Get features of Korean input texts w/ BM-K/KoSimCSE-roberta.
+    Returns:
+        List[List[int]] of dimension 768
+    """
+
+    model = AutoModel.from_pretrained("BM-K/KoSimCSE-roberta")
+    tokenizer = AutoTokenizer.from_pretrained("BM-K/KoSimCSE-roberta")
+    inputs = tokenizer(sentences, padding=True, truncation=True, return_tensors="pt")
+    embeddings, _ = model(**inputs, return_dict=False)
+    ls = []
+    for embedding in embeddings:
+        vector = embedding[0].detach().numpy().tolist()
+        ls.append(vector)
+    return ls
+
+
+class KorRobertaEmbeddings(Embeddings):
+    """Feature Extraction w/ BM-K/KoSimCSE-roberta"""
+
+    dimension = 768
+
+    def embed_documents(self, texts: List[str]) -> List[List[float]]:
+        return get_roberta_embeddings(texts)
+
+    def embed_query(self, text: str) -> List[float]:
+        return get_roberta_embeddings([text])[0]

requirements.txt

@@ -0,0 +1,83 @@
+-i https://pypi.org/simple
+aiohttp==3.9.3; python_version >= '3.8'
+aiosignal==1.3.1; python_version >= '3.7'
+altair==5.2.0; python_version >= '3.8'
+annotated-types==0.6.0; python_version >= '3.8'
+anyio==4.3.0; python_version >= '3.8'
+async-timeout==4.0.3; python_version < '3.11'
+attrs==23.2.0; python_version >= '3.7'
+blinker==1.7.0; python_version >= '3.8'
+cachetools==5.3.2; python_version >= '3.7'
+certifi==2024.2.2; python_version >= '3.6'
+charset-normalizer==3.3.2; python_full_version >= '3.7.0'
+click==8.1.7; python_version >= '3.7'
+dataclasses-json==0.6.4; python_version >= '3.7' and python_version < '4.0'
+exceptiongroup==1.2.0; python_version < '3.11'
+filelock==3.13.1; python_version >= '3.8'
+frozenlist==1.4.1; python_version >= '3.8'
+fsspec==2024.2.0; python_version >= '3.8'
+gitdb==4.0.11; python_version >= '3.7'
+gitpython==3.1.42; python_version >= '3.7'
+huggingface-hub==0.20.3; python_full_version >= '3.8.0'
+idna==3.6; python_version >= '3.5'
+importlib-metadata==7.0.1; python_version >= '3.8'
+jinja2==3.1.3; python_version >= '3.7'
+jsonpatch==1.33; python_version >= '2.7' and python_version not in '3.0, 3.1, 3.2, 3.3, 3.4, 3.5, 3.6'
+jsonpointer==2.4; python_version >= '2.7' and python_version not in '3.0, 3.1, 3.2, 3.3, 3.4, 3.5, 3.6'
+jsonschema==4.21.1; python_version >= '3.8'
+jsonschema-specifications==2023.12.1; python_version >= '3.8'
+langchain==0.1.8; python_version < '4.0' and python_full_version >= '3.8.1'
+langchain-community==0.0.21; python_version < '4.0' and python_full_version >= '3.8.1'
+langchain-core==0.1.25; python_version < '4.0' and python_full_version >= '3.8.1'
+langchain-pinecone==0.0.2; python_version < '3.13' and python_full_version >= '3.8.1'
+langsmith==0.1.5; python_version < '4.0' and python_full_version >= '3.8.1'
+markdown-it-py==3.0.0; python_version >= '3.8'
+markupsafe==2.1.5; python_version >= '3.7'
+marshmallow==3.20.2; python_version >= '3.8'
+mdurl==0.1.2; python_version >= '3.7'
+mpmath==1.3.0
+multidict==6.0.5; python_version >= '3.7'
+mypy-extensions==1.0.0; python_version >= '3.5'
+networkx==3.2.1; python_version >= '3.9'
+numpy==1.26.4; python_version >= '3.9'
+packaging==23.2; python_version >= '3.7'
+pandas==2.2.0; python_version >= '3.9'
+pillow==10.2.0; python_version >= '3.8'
+pinecone-client==3.0.3; python_version < '3.13' and python_version >= '3.8'
+protobuf==4.25.3; python_version >= '3.8'
+pyarrow==15.0.0; python_version >= '3.8'
+pydantic==2.6.1; python_version >= '3.8'
+pydantic-core==2.16.2; python_version >= '3.8'
+pydeck==0.8.1b0; python_version >= '3.7'
+pygments==2.17.2; python_version >= '3.7'
+python-dateutil==2.8.2; python_version >= '2.7' and python_version not in '3.0, 3.1, 3.2, 3.3'
+pytz==2024.1
+pyyaml==6.0.1; python_version >= '3.6'
+referencing==0.33.0; python_version >= '3.8'
+regex==2023.12.25; python_version >= '3.7'
+requests==2.31.0; python_version >= '3.7'
+rich==13.7.0; python_full_version >= '3.7.0'
+rpds-py==0.18.0; python_version >= '3.8'
+safetensors==0.4.2; python_version >= '3.7'
+six==1.16.0; python_version >= '2.7' and python_version not in '3.0, 3.1, 3.2, 3.3'
+smmap==5.0.1; python_version >= '3.7'
+sniffio==1.3.0; python_version >= '3.7'
+sqlalchemy==2.0.27; python_version >= '3.7'
+streamlit==1.31.1; python_version >= '3.8' and python_full_version != '3.9.7'
+sympy==1.12; python_version >= '3.8'
+tenacity==8.2.3; python_version >= '3.7'
+tokenizers==0.15.2; python_version >= '3.7'
+toml==0.10.2; python_version >= '2.6' and python_version not in '3.0, 3.1, 3.2, 3.3'
+toolz==0.12.1; python_version >= '3.7'
+torch==2.2.0; python_full_version >= '3.8.0'
+tornado==6.4; python_version >= '3.8'
+tqdm==4.66.2; python_version >= '3.7'
+transformers==4.37.2; python_full_version >= '3.8.0'
+typing-extensions==4.9.0; python_version >= '3.8'
+typing-inspect==0.9.0
+tzdata==2024.1; python_version >= '2'
+tzlocal==5.2; python_version >= '3.8'
+urllib3==2.2.1; python_version >= '3.8'
+validators==0.22.0; python_version >= '3.8'
+yarl==1.9.4; python_version >= '3.7'
+zipp==3.17.0; python_version >= '3.8'