|
import os |
|
import torch |
|
from torch import cuda, bfloat16 |
|
from transformers import AutoTokenizer, pipeline, BitsAndBytesConfig |
|
from langchain.llms import HuggingFacePipeline |
|
from langchain.vectorstores import FAISS |
|
from langchain.chains import ConversationalRetrievalChain |
|
import gradio as gr |
|
from langchain.embeddings import HuggingFaceEmbeddings |
|
from transformers import InferenceClient |
|
|
|
|
|
HF_TOKEN = os.environ.get("HF_TOKEN", None) |
|
|
|
|
|
model_id = 'mistralai/Mistral-7B-Instruct-v0.3' |
|
client = InferenceClient(model_id) |
|
|
|
|
|
class StopOnTokens: |
|
def __call__(self, input_ids, scores, **kwargs): |
|
for stop_ids in stop_token_ids: |
|
if torch.eq(input_ids[0][-len(stop_ids):], stop_ids).all(): |
|
return True |
|
return False |
|
|
|
|
|
stop_list = ['\nHuman:', '\n```\n'] |
|
stop_token_ids = [client.tokenizer(x)['input_ids'] for x in stop_list] |
|
stop_token_ids = [torch.LongTensor(x).to(cuda.current_device() if cuda.is_available() else 'cpu') for x in stop_token_ids] |
|
|
|
|
|
def generate(prompt, history, system_prompt=None, temperature=0.2, max_new_tokens=1024, top_p=0.95, repetition_penalty=1.0): |
|
temperature = float(temperature) |
|
if temperature < 1e-2: |
|
temperature = 1e-2 |
|
top_p = float(top_p) |
|
|
|
generate_kwargs = dict( |
|
temperature=temperature, |
|
max_new_tokens=max_new_tokens, |
|
top_p=top_p, |
|
repetition_penalty=repetition_penalty, |
|
do_sample=True, |
|
seed=42, |
|
) |
|
|
|
formatted_prompt = format_prompt(prompt, history, system_prompt) |
|
stream = client.text_generation(formatted_prompt, **generate_kwargs, stream=True, details=True, return_full_text=False) |
|
output = "" |
|
|
|
for response in stream: |
|
output += response.token.text |
|
yield output |
|
return output |
|
|
|
llm = HuggingFacePipeline(pipeline=generate) |
|
|
|
|
|
try: |
|
vectorstore = FAISS.load_local('faiss_index', HuggingFaceEmbeddings(model_name="sentence-transformers/all-mpnet-base-v2", model_kwargs={"device": "cuda"})) |
|
print("Loaded embedding successfully") |
|
except ImportError as e: |
|
print("FAISS could not be imported. Make sure FAISS is installed correctly.") |
|
raise e |
|
|
|
|
|
chain = ConversationalRetrievalChain.from_llm(llm, vectorstore.as_retriever(), return_source_documents=True) |
|
|
|
chat_history = [] |
|
|
|
def format_prompt(query): |
|
prompt = f""" |
|
You are a knowledgeable assistant with access to a comprehensive database. |
|
I need you to answer my question and provide related information in a specific format. |
|
Here's what I need: |
|
1. A brief, general response to my question based on related answers retrieved. |
|
2. A JSON-formatted output containing: |
|
- "question": The original question. |
|
- "answer": The detailed answer. |
|
- "related_questions": A list of related questions and their answers, each as a dictionary with the keys: |
|
- "question": The related question. |
|
- "answer": The related answer. |
|
Here's my question: |
|
{query} |
|
Include a brief final answer without additional comments, sign-offs, or extra phrases. Be direct and to the point. |
|
""" |
|
return prompt |
|
|
|
def qa_infer(query): |
|
formatted_prompt = format_prompt(query) |
|
result = chain({"question": formatted_prompt, "chat_history": chat_history}) |
|
for doc in result['source_documents']: |
|
print("-"*50) |
|
print("Retrieved Document:", doc.page_content) |
|
print("#"*100) |
|
print(result['answer']) |
|
return result['answer'] |
|
|
|
EXAMPLES = ["How to use IPU1_0 instead of A15_0 to process NDK in TDA2x-EVM", |
|
"Can BQ25896 support I2C interface?", |
|
"Does TDA2 vout support bt656 8-bit mode?"] |
|
|
|
demo = gr.Interface(fn=qa_infer, inputs="text", allow_flagging='never', examples=EXAMPLES, cache_examples=False, outputs="text") |
|
demo.launch() |
|
|