2023-10-28 23:47:18 +02:00

87 lines
2.6 KiB
Python

import gradio as gr
import sys
import os
from collections.abc import Iterable
from langchain.document_loaders import PyPDFLoader, Docx2txtLoader, TextLoader, UnstructuredHTMLLoader
from langchain.text_splitter import RecursiveCharacterTextSplitter, CharacterTextSplitter
from langchain.chains import RetrievalQA
from langchain.llms import Ollama
from langchain.vectorstores import FAISS, Chroma
from langchain.embeddings import GPT4AllEmbeddings, CacheBackedEmbeddings
from langchain.storage import LocalFileStore#, RedisStore, UpstashRedisStore, InMemoryStore
ollama = Ollama(base_url='http://localhost:11434',
#model="codellama")
#model="starcoder")
model="llama2")
docsUrl = "/home/user/dev/docs"
documents = []
for file in os.listdir(docsUrl):
if file.endswith(".pdf"):
pdf_path = docsUrl + "/" + file
loader = PyPDFLoader(pdf_path)
documents.extend(loader.load())
print("Found " + pdf_path)
elif file.endswith('.docx') or file.endswith('.doc'):
doc_path = docsUrl + "/" + file
loader = Docx2txtLoader(doc_path)
documents.extend(loader.load())
print("Found " + doc_path)
elif file.endswith('.txt') or file.endswith('.kt') or file.endswith('.json'):
text_path = docsUrl + "/" + file
loader = TextLoader(text_path)
documents.extend(loader.load())
print("Found " + text_path)
elif file.endswith('.html') or file.endswith('.htm'):
htm_path = docsUrl + "/" + file
loader = UnstructuredHTMLLoader(htm_path)
documents.extend(loader.load())
print("Found " + htm_path)
text_splitter = CharacterTextSplitter(chunk_size=32, chunk_overlap=32)
all_splits = text_splitter.split_documents(documents)
#fs = LocalFileStore("/home/gabriele/dev/cache/")
#underlying_embeddings = GPT4AllEmbeddings()
#cached_embedder = CacheBackedEmbeddings.from_bytes_store(
# underlying_embeddings, fs, namespace=underlying_embeddings.model
#)
vectorstore = Chroma.from_documents(documents=all_splits, embedding=GPT4AllEmbeddings(embeddings_chunk_size=1000))
#vectorstore = FAISS.from_documents(documents=all_splits, embedding=cached_embedder)
def AI_response(question, history):
docs = vectorstore.similarity_search(question)
len(docs)
qachain=RetrievalQA.from_chain_type(ollama, retriever=vectorstore.as_retriever())
#reply=qachain()
#reply=str(qachain({"query": question}))
reply=str(qachain.run(question))
return reply
demo = gr.ChatInterface(AI_response, title="Put your files in folder" + docsUrl)
if __name__ == "__main__":
demo.launch(server_name="0.0.0.0", server_port=7860)