app.py

import gradio as gr

from langchain.embeddings import HuggingFaceEmbeddings, HuggingFaceInstructEmbeddings, OpenAIEmbeddings
from langchain.vectorstores import Pinecone
import pinecone
import os
os.environ["TOKENIZERS_PARALLELISM"] = "false"


PINECONE_KEY = os.environ.get("PINECONE_KEY", "")
PINECONE_ENV = os.environ.get("PINECONE_ENV", "asia-northeast1-gcp")
PINECONE_INDEX = os.environ.get("PINECONE_INDEX", '3gpp-r16')

EMBEDDING_MODEL = os.environ.get("EMBEDDING_MODEL", "hkunlp/instructor-large")
EMBEDDING_LOADER = os.environ.get("EMBEDDING_LOADER", "HuggingFaceInstructEmbeddings")
EMBEDDING_LIST = ["HuggingFaceInstructEmbeddings", "HuggingFaceEmbeddings"]

# return top-k text chunks from vector store
TOP_K_DEFAULT = 15
TOP_K_MAX = 30
SCORE_DEFAULT = 0.33

global g_db
g_db = None

def init_db(emb_name, emb_loader, db_api_key, db_env, db_index):

    embeddings = eval(emb_loader)(model_name=emb_name)

    pinecone.init(api_key     = db_api_key,
                  environment = db_env)

    global g_db

    g_db = Pinecone.from_existing_index(index_name = db_index,
                                      embedding  = embeddings)
    return str(g_db)


def get_db():
    return g_db


def remove_duplicates(documents, score_min):
    seen_content = set()
    unique_documents = []
    for (doc, score) in documents:
        if (doc.page_content not in seen_content) and (score >= score_min):
            seen_content.add(doc.page_content)
            unique_documents.append(doc)
    return unique_documents


def get_data(query, top_k, score):
    if not query:
        return "Please init db in configuration"

    print("Use db: " + str(g_db))

    docs = g_db.similarity_search_with_score(query = query,
                                             k=top_k)
    #docsearch = db.as_retriever(search_kwargs={'k':top_k})
    #docs = docsearch.get_relevant_documents(query)
    udocs = remove_duplicates(docs, score)
    return udocs

with gr.Blocks(
    title = "3GPP Database",
    theme = "Base",
    css = """.bigbox {
    min-height:250px;
}
""") as demo:
    with gr.Tab("Matching"):
        with gr.Accordion("Vector similarity"):
            with gr.Row():
                with gr.Column():
                    top_k = gr.Slider(1,
                                      TOP_K_MAX,
                                      value=TOP_K_DEFAULT,
                                      step=1,
                                      label="Vector similarity top_k",
                                      interactive=True)
                with gr.Column():
                    score = gr.Slider(0.01,
                                      0.99,
                                      value=SCORE_DEFAULT,
                                      step=0.01,
                                      label="Vector similarity score",
                                      interactive=True)

        with gr.Row():
             inp = gr.Textbox(label = "Input",
                              placeholder="What are you looking for?")
             out = gr.Textbox(label = "Output")

        btn_run = gr.Button("Run", variant="primary")

    with gr.Tab("Configuration"):
        with gr.Row():
            loading = gr.Textbox(get_db, max_lines=1, show_label=False)
            btn_init = gr.Button("Init")
        with gr.Accordion("Embedding"):
            with gr.Row():
                with gr.Column():
                    emb_textbox = gr.Textbox(
                        label = "Embedding Model",
                        # show_label = False,
                        value = EMBEDDING_MODEL,
                        placeholder = "Paste Your Embedding Model Repo on HuggingFace",
                        lines=1,
                        interactive=True,
                        type='email')

                with gr.Column():
                    emb_dropdown = gr.Dropdown(
                        EMBEDDING_LIST,
                        value=EMBEDDING_LOADER,
                        multiselect=False,
                        interactive=True,
                        label="Embedding Loader")

        with gr.Accordion("Pinecone Database"):
            with gr.Row():
                db_api_textbox = gr.Textbox(
                    label = "Pinecone API Key",
                    # show_label = False,
                    value = PINECONE_KEY,
                    placeholder = "Paste Your Pinecone API Key (xx-xx-xx-xx-xx) and Hit ENTER",
                    lines=1,
                    interactive=True,
                    type='password')
            with gr.Row():
                db_env_textbox = gr.Textbox(
                    label = "Pinecone Environment",
                    # show_label = False,
                    value = PINECONE_ENV,
                    placeholder = "Paste Your Pinecone Environment (xx-xx-xx) and Hit ENTER",
                    lines=1,
                    interactive=True,
                    type='email')
                db_index_textbox = gr.Textbox(
                    label = "Pinecone Index",
                    # show_label = False,
                    value = PINECONE_INDEX,
                    placeholder = "Paste Your Pinecone Index (xxxx) and Hit ENTER",
                    lines=1,
                    interactive=True,
                    type='email')

    btn_init.click(fn=init_db, inputs=[emb_textbox, emb_dropdown, db_api_textbox, db_env_textbox, db_index_textbox], outputs=loading)
    btn_run.click(fn=get_data, inputs=[inp, top_k, score], outputs=out)

if __name__ == "__main__":
    demo.queue()
    demo.launch(inbrowser = True)