Spaces:
Sleeping
Sleeping
arjunanand13
commited on
Commit
•
2497fee
1
Parent(s):
eae3c36
Create app.py
Browse files
app.py
ADDED
@@ -0,0 +1,105 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import os
|
2 |
+
import torch
|
3 |
+
from torch import cuda, bfloat16
|
4 |
+
from transformers import AutoTokenizer, pipeline, BitsAndBytesConfig
|
5 |
+
from langchain.llms import HuggingFacePipeline
|
6 |
+
from langchain.vectorstores import FAISS
|
7 |
+
from langchain.chains import ConversationalRetrievalChain
|
8 |
+
import gradio as gr
|
9 |
+
from langchain.embeddings import HuggingFaceEmbeddings
|
10 |
+
from transformers import InferenceClient
|
11 |
+
|
12 |
+
# Load the Hugging Face token from environment
|
13 |
+
HF_TOKEN = os.environ.get("HF_TOKEN", None)
|
14 |
+
|
15 |
+
# Load the Mistral model and tokenizer
|
16 |
+
model_id = 'mistralai/Mistral-7B-Instruct-v0.3'
|
17 |
+
client = InferenceClient(model_id)
|
18 |
+
|
19 |
+
# Define stopping criteria
|
20 |
+
class StopOnTokens:
|
21 |
+
def __call__(self, input_ids, scores, **kwargs):
|
22 |
+
for stop_ids in stop_token_ids:
|
23 |
+
if torch.eq(input_ids[0][-len(stop_ids):], stop_ids).all():
|
24 |
+
return True
|
25 |
+
return False
|
26 |
+
|
27 |
+
# Define stopping criteria list
|
28 |
+
stop_list = ['\nHuman:', '\n```\n']
|
29 |
+
stop_token_ids = [client.tokenizer(x)['input_ids'] for x in stop_list]
|
30 |
+
stop_token_ids = [torch.LongTensor(x).to(cuda.current_device() if cuda.is_available() else 'cpu') for x in stop_token_ids]
|
31 |
+
|
32 |
+
# Create text generation pipeline
|
33 |
+
def generate(prompt, history, system_prompt=None, temperature=0.2, max_new_tokens=1024, top_p=0.95, repetition_penalty=1.0):
|
34 |
+
temperature = float(temperature)
|
35 |
+
if temperature < 1e-2:
|
36 |
+
temperature = 1e-2
|
37 |
+
top_p = float(top_p)
|
38 |
+
|
39 |
+
generate_kwargs = dict(
|
40 |
+
temperature=temperature,
|
41 |
+
max_new_tokens=max_new_tokens,
|
42 |
+
top_p=top_p,
|
43 |
+
repetition_penalty=repetition_penalty,
|
44 |
+
do_sample=True,
|
45 |
+
seed=42,
|
46 |
+
)
|
47 |
+
|
48 |
+
formatted_prompt = format_prompt(prompt, history, system_prompt)
|
49 |
+
stream = client.text_generation(formatted_prompt, **generate_kwargs, stream=True, details=True, return_full_text=False)
|
50 |
+
output = ""
|
51 |
+
|
52 |
+
for response in stream:
|
53 |
+
output += response.token.text
|
54 |
+
yield output
|
55 |
+
return output
|
56 |
+
|
57 |
+
llm = HuggingFacePipeline(pipeline=generate)
|
58 |
+
|
59 |
+
# Load the stored FAISS index
|
60 |
+
try:
|
61 |
+
vectorstore = FAISS.load_local('faiss_index', HuggingFaceEmbeddings(model_name="sentence-transformers/all-mpnet-base-v2", model_kwargs={"device": "cuda"}))
|
62 |
+
print("Loaded embedding successfully")
|
63 |
+
except ImportError as e:
|
64 |
+
print("FAISS could not be imported. Make sure FAISS is installed correctly.")
|
65 |
+
raise e
|
66 |
+
|
67 |
+
# Set up the Conversational Retrieval Chain
|
68 |
+
chain = ConversationalRetrievalChain.from_llm(llm, vectorstore.as_retriever(), return_source_documents=True)
|
69 |
+
|
70 |
+
chat_history = []
|
71 |
+
|
72 |
+
def format_prompt(query):
|
73 |
+
prompt = f"""
|
74 |
+
You are a knowledgeable assistant with access to a comprehensive database.
|
75 |
+
I need you to answer my question and provide related information in a specific format.
|
76 |
+
Here's what I need:
|
77 |
+
1. A brief, general response to my question based on related answers retrieved.
|
78 |
+
2. A JSON-formatted output containing:
|
79 |
+
- "question": The original question.
|
80 |
+
- "answer": The detailed answer.
|
81 |
+
- "related_questions": A list of related questions and their answers, each as a dictionary with the keys:
|
82 |
+
- "question": The related question.
|
83 |
+
- "answer": The related answer.
|
84 |
+
Here's my question:
|
85 |
+
{query}
|
86 |
+
Include a brief final answer without additional comments, sign-offs, or extra phrases. Be direct and to the point.
|
87 |
+
"""
|
88 |
+
return prompt
|
89 |
+
|
90 |
+
def qa_infer(query):
|
91 |
+
formatted_prompt = format_prompt(query)
|
92 |
+
result = chain({"question": formatted_prompt, "chat_history": chat_history})
|
93 |
+
for doc in result['source_documents']:
|
94 |
+
print("-"*50)
|
95 |
+
print("Retrieved Document:", doc.page_content)
|
96 |
+
print("#"*100)
|
97 |
+
print(result['answer'])
|
98 |
+
return result['answer']
|
99 |
+
|
100 |
+
EXAMPLES = ["How to use IPU1_0 instead of A15_0 to process NDK in TDA2x-EVM",
|
101 |
+
"Can BQ25896 support I2C interface?",
|
102 |
+
"Does TDA2 vout support bt656 8-bit mode?"]
|
103 |
+
|
104 |
+
demo = gr.Interface(fn=qa_infer, inputs="text", allow_flagging='never', examples=EXAMPLES, cache_examples=False, outputs="text")
|
105 |
+
demo.launch()
|