-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathLCEL.py
268 lines (228 loc) · 11.3 KB
/
LCEL.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
import pandas as pd
from langchain.prompts import PromptTemplate
from langchain.schema import SystemMessage, HumanMessage
from langchain.memory import ConversationBufferMemory
import streamlit as st
from langchain_core.runnables import RunnablePassthrough, RunnableLambda
from langchain_core.pydantic_v1 import BaseModel, Field
from langchain_core.output_parsers import StrOutputParser, JsonOutputParser
from langchain_core.documents import Document
from operator import itemgetter
from langchain_core.messages import get_buffer_string
from langchain.schema import format_document
from pymilvus.model.reranker import CrossEncoderRerankFunction
from config import compression_retriever, user_compression_retriever, db_doc_k, rerank_model_name
from prompt import not_rag_template, rag_template, llama_template, img_rag_template, PROMPT_INJECTION_PROMPT, SYSTEMPROMPT
from LLM import HCX, gpt_model, sonnet_llm, sllm, gemini_vis_model, gemini_txt_model
ONLY_CHAIN_PROMPT = PromptTemplate(input_variables=["question"],template=not_rag_template)
QA_CHAIN_PROMPT = PromptTemplate(input_variables=["context", "question"],template=rag_template)
LLAMA_QA_CHAIN_PROMPT = PromptTemplate(input_variables=["context", "question"],template=llama_template)
IMG_QA_CHAIN_PROMPT = PromptTemplate(input_variables=["img_context", "valid_img_context", "question"],template=img_rag_template)
class doc_review_system_key(BaseModel):
문서명: str = Field(description="문서명을 추출해줘.")
조합명: str = Field(description="조합명(단체명)을 추출해줘, 단 추출 시, 띄어쓰기를 반드시 주의해야 해.")
고유번호: str = Field(description="고유번호를 추출해줘, 숫자 3자리 - 숫자 2자리 - 숫자 5자리 형태야.")
주소: str = Field(description="주소(소재지)를 추출해줘.")
대표자성명: str = Field(description="대표자 성명을 추출해줘.")
생년월일: str = Field(description="생년월일을 추출해줘.")
조합인감: str = Field(description="조합인감의 유/무를 O/X로 나타내줘.")
json_parser = JsonOutputParser(pydantic_object=doc_review_system_key)
def multimodal_prompt_json_parser(img_base64):
return [
SystemMessage(
content=""""Answer the user question.\n{format_instructions}\n{question}\n"""
),
HumanMessage(
content=[
{
"type": "image_url",
"image_url": {
"url": f"data:image/png;base64,{img_base64}",
},
},
{
"type": "text", "text": json_parser.get_format_instructions()
}
]
)
]
hcx_sec = HCX(init_system_prompt = PROMPT_INJECTION_PROMPT)
hcx_stream = HCX(init_system_prompt = SYSTEMPROMPT)
@st.cache_resource
def asa_init_memory():
return ConversationBufferMemory(
input_key='question',
output_key='answer',
memory_key='chat_history',
return_messages=True)
asa_memory = asa_init_memory()
@st.cache_resource
def hcx_init_memory():
return ConversationBufferMemory(
input_key='question',
output_key='answer',
memory_key='chat_history',
return_messages=True)
hcx_memory = hcx_init_memory()
@st.cache_resource
def gpt_init_memory():
return ConversationBufferMemory(
input_key='question',
output_key='answer',
memory_key='chat_history',
return_messages=True)
gpt_memory = gpt_init_memory()
@st.cache_resource
def sllm_init_memory():
return ConversationBufferMemory(
input_key='question',
output_key='answer',
memory_key='chat_history',
return_messages=True)
sllm_memory = sllm_init_memory()
@st.cache_resource
def gemini_init_memory():
return ConversationBufferMemory(
input_key='question',
output_key='answer',
memory_key='chat_history',
return_messages=True)
gemini_memory = gemini_init_memory()
# reset button
def reset_conversation():
st.session_state.conversation = None
st.session_state.chat_history = None
asa_memory.clear()
hcx_memory.clear()
gpt_memory.clear()
sllm_memory.clear()
gemini_memory.clear()
st.toast("대화 리셋!", icon="👋")
DEFAULT_DOCUMENT_PROMPT = PromptTemplate.from_template(template="{page_content}")
def _combine_documents(
docs, document_prompt=DEFAULT_DOCUMENT_PROMPT, document_separator="\n\n"
):
doc_strings = [format_document(doc, document_prompt) for doc in docs]
return document_separator.join(doc_strings)
asa_loaded_memory = RunnablePassthrough.assign(
chat_history=RunnableLambda(asa_memory.load_memory_variables) | itemgetter("chat_history"),
)
hcx_loaded_memory = RunnablePassthrough.assign(
chat_history=RunnableLambda(hcx_memory.load_memory_variables) | itemgetter("chat_history"),
)
gpt_loaded_memory = RunnablePassthrough.assign(
chat_history=RunnableLambda(gpt_memory.load_memory_variables) | itemgetter("chat_history"),
)
sllm_loaded_memory = RunnablePassthrough.assign(
chat_history=RunnableLambda(sllm_memory.load_memory_variables) | itemgetter("chat_history"),
)
gemini_loaded_memory = RunnablePassthrough.assign(
chat_history=RunnableLambda(gemini_memory.load_memory_variables) | itemgetter("chat_history"),
)
retrieved_documents = {
"question": lambda x: x["question"],
"pre_context": itemgetter("question") | compression_retriever,
"chat_history": lambda x: get_buffer_string(x["chat_history"])
}
user_retrieved_documents = {
"question": lambda x: x["question"],
"pre_context": itemgetter("question") | user_compression_retriever,
"chat_history": lambda x: get_buffer_string(x["chat_history"])
}
@st.cache_resource
def init_rerank_model():
return CrossEncoderRerankFunction(
model_name=rerank_model_name,
device="cpu"
)
rerank_model = init_rerank_model()
class SrcDoc:
post_src_doc: list
formatted_metadata: list
def src_doc(self, init_src_doc):
if len(init_src_doc['pre_context']) > 0:
source_documents_total = init_src_doc['pre_context']
src_doc_context = [doc.page_content for doc in source_documents_total]
src_doc_score = [doc.state['query_similarity_score'] for doc in source_documents_total]
src_doc_metadata = [doc.metadata['source'] for doc in source_documents_total]
src_doc_page = [doc.metadata['page'] for doc in source_documents_total]
src_doc_ori_index = [i for i in range(len(source_documents_total))]
src_doc_ori_index_doc_page = [src_doc_ori_index, src_doc_metadata, src_doc_page]
src_doc_ori_index_doc_page = list(zip(*src_doc_ori_index_doc_page))
# 통합된문서 기반 rerank 진행
ce_rerank_docs = rerank_model(init_src_doc["question"], src_doc_context, top_k = db_doc_k)
ce_rerank_text = [i.text for i in ce_rerank_docs]
ce_rerank_index = [i.index for i in ce_rerank_docs]
ce_rerank_metadata = [src_doc_ori_index_doc_page[i][1] for i in ce_rerank_index]
ce_rerank_page = [src_doc_ori_index_doc_page[i][2] for i in ce_rerank_index]
self.formatted_metadata = [
f'문서 명: {metadata.split("/")[-1]}, 문서 위치: {page} 쪽'
for metadata, page in zip(ce_rerank_metadata, ce_rerank_page)
]
src_doc_df = pd.DataFrame({'내용': ce_rerank_text, '문서 출처': self.formatted_metadata})
src_doc_df['No'] = [i+1 for i in range(src_doc_df.shape[0])]
src_doc_df = src_doc_df.set_index('No')
src_doc_df['내용'] = src_doc_df['내용'].str.slice(0, 100) + '.....(이하 생략)'
with st.expander('참조 문서'):
st.table(src_doc_df)
# ce_rerank_docs의 langchain document 화
self.post_src_doc= [
Document(
page_content=text,
metadata={"source": metadata}
)
for text, metadata in zip(ce_rerank_text, ce_rerank_metadata)
]
post_src_doc_output = {"question": init_src_doc["question"], "post_context": self.post_src_doc, "chat_history": init_src_doc["chat_history"]}
else:
src_doc_context = []
src_doc_metadata = []
self.post_src_doc= [
Document(
page_content=text,
metadata={"source": metadata}
)
for text, metadata in zip(src_doc_context, src_doc_metadata)
]
post_src_doc_output = {"question": init_src_doc["question"], "post_context": self.post_src_doc, "chat_history": init_src_doc["chat_history"]}
return post_src_doc_output
src_doc = SrcDoc()
not_retrieved_documents = {
"question": lambda x: x["question"],
"chat_history": lambda x: get_buffer_string(x["chat_history"])
}
img_retrieved_documents = {
"question": lambda x: x["question"],
"img_context": lambda x: x["img_context"],
"valid_img_context": lambda x: x["valid_img_context"],
"chat_history": lambda x: get_buffer_string(x["chat_history"])
}
final_inputs = {
"context": lambda x: _combine_documents(x["post_context"]),
"question": itemgetter("question"),
"chat_history": itemgetter("chat_history")
}
img_final_inputs = {
"img_context": itemgetter("img_context"),
"valid_img_context": itemgetter("valid_img_context"),
"question": itemgetter("question"),
"chat_history": itemgetter("chat_history")
}
hcx_sec_pipe = ONLY_CHAIN_PROMPT | hcx_sec | StrOutputParser()
retrieval_qa_chain = asa_loaded_memory | retrieved_documents | (lambda x: src_doc.src_doc(x)) | final_inputs | QA_CHAIN_PROMPT | hcx_stream | StrOutputParser()
user_retrieval_qa_chain = asa_loaded_memory | user_retrieved_documents | (lambda x: src_doc.src_doc(x)) | final_inputs | QA_CHAIN_PROMPT | hcx_stream | StrOutputParser()
hcx_only_pipe = hcx_loaded_memory | not_retrieved_documents | ONLY_CHAIN_PROMPT | hcx_stream | StrOutputParser()
gpt_pipe = gpt_loaded_memory | not_retrieved_documents | ONLY_CHAIN_PROMPT | gpt_model | StrOutputParser()
aws_retrieval_qa_chain = asa_loaded_memory | retrieved_documents | (lambda x: src_doc.src_doc(x)) | final_inputs | QA_CHAIN_PROMPT | sonnet_llm | StrOutputParser()
sllm_pipe = sllm_loaded_memory | retrieved_documents | (lambda x: src_doc.src_doc(x)) | final_inputs | LLAMA_QA_CHAIN_PROMPT | sllm | StrOutputParser()
gemini_txt_pipe = gemini_loaded_memory | not_retrieved_documents | ONLY_CHAIN_PROMPT | gemini_txt_model | StrOutputParser()
gemini_vis_pipe = RunnablePassthrough() | gemini_vis_model | StrOutputParser()
gemini_vis_txt_pipe = (
gemini_loaded_memory | img_retrieved_documents | img_final_inputs |
IMG_QA_CHAIN_PROMPT | gemini_txt_model | StrOutputParser()
)
aws_vis_json_parser_pipe = RunnablePassthrough() | sonnet_llm | json_parser
aws_vis_txt_pipe = (
asa_loaded_memory | img_retrieved_documents | img_final_inputs |
IMG_QA_CHAIN_PROMPT | sonnet_llm | StrOutputParser()
)