Spaces:
Sleeping
Sleeping
Update azure_openai.py
Browse files- azure_openai.py +348 -348
azure_openai.py
CHANGED
@@ -1,349 +1,349 @@
|
|
1 |
-
import streamlit as st
|
2 |
-
import os
|
3 |
-
import pandas as pd
|
4 |
-
# from langchain.chat_models import AzureChatOpenAI
|
5 |
-
from langchain_openai import AzureChatOpenAI
|
6 |
-
from langchain_core.output_parsers import StrOutputParser, PydanticOutputParser
|
7 |
-
from langchain_core.prompts.chat import SystemMessagePromptTemplate, HumanMessagePromptTemplate, ChatPromptTemplate
|
8 |
-
from pydantic import BaseModel, Field, validator
|
9 |
-
from langchain.output_parsers.enum import EnumOutputParser
|
10 |
-
from langchain_core.prompts import PromptTemplate
|
11 |
-
from enum import Enum
|
12 |
-
|
13 |
-
|
14 |
-
os.environ["LANGCHAIN_TRACING_V2"]="true"
|
15 |
-
os.environ["LANGCHAIN_ENDPOINT"]="https://api.smith.langchain.com"
|
16 |
-
LANGCHAIN_API_KEY = st.secrets['LANGCHAIN_API_KEY']
|
17 |
-
os.environ["LANGCHAIN_PROJECT"]="UC2e2e"
|
18 |
-
|
19 |
-
# LLM Langchain Definition
|
20 |
-
OPENAI_API_KEY = st.secrets['OPENAI_API_KEY']
|
21 |
-
OPENAI_API_TYPE = "azure"
|
22 |
-
OPENAI_API_BASE = "https://davidfearn-gpt4.openai.azure.com"
|
23 |
-
OPENAI_API_VERSION = "2024-08-01-preview"
|
24 |
-
OPENAI_MODEL = "gpt-4o-mini"
|
25 |
-
|
26 |
-
|
27 |
-
# Function to read file contents
|
28 |
-
def read_file(file):
|
29 |
-
"""
|
30 |
-
Reads the content of a text file and returns it as a string.
|
31 |
-
:param file: The file name to read from the 'assets' directory.
|
32 |
-
:return: The content of the file as a string or None if an error occurs.
|
33 |
-
"""
|
34 |
-
fp = f"assets/{file}.md"
|
35 |
-
try:
|
36 |
-
with open(fp, 'r', encoding='utf-8') as file:
|
37 |
-
content = file.read()
|
38 |
-
return content
|
39 |
-
except FileNotFoundError:
|
40 |
-
print(f"The file at {fp} was not found.")
|
41 |
-
except IOError:
|
42 |
-
print(f"An error occurred while reading the file at {fp}.")
|
43 |
-
return None
|
44 |
-
|
45 |
-
# Function to generate structured insights
|
46 |
-
def process_insight(chunk, topic,source):
|
47 |
-
|
48 |
-
GSKGlossary = read_file("GSKGlossary")
|
49 |
-
if source== "intl":
|
50 |
-
SystemMessage = read_file("intl_insight_system_message")
|
51 |
-
UserMessage = read_file("intl_insight_user_message")
|
52 |
-
else:
|
53 |
-
SystemMessage = read_file("ext_insight_system_message")
|
54 |
-
UserMessage = read_file("ext_insight_user_message")
|
55 |
-
|
56 |
-
|
57 |
-
class Insights(BaseModel):
|
58 |
-
completed: bool = Field(description="This field is used to indicate that you think the number of insights has been completed")
|
59 |
-
insight: str = Field(description="This field is used to return the MECE insight in string format")
|
60 |
-
|
61 |
-
|
62 |
-
llm = AzureChatOpenAI(
|
63 |
-
openai_api_version=OPENAI_API_VERSION,
|
64 |
-
openai_api_key=OPENAI_API_KEY,
|
65 |
-
azure_endpoint=OPENAI_API_BASE,
|
66 |
-
openai_api_type=OPENAI_API_TYPE,
|
67 |
-
deployment_name=OPENAI_MODEL,
|
68 |
-
temperature=0,
|
69 |
-
)
|
70 |
-
|
71 |
-
system_message_template = SystemMessagePromptTemplate.from_template(SystemMessage)
|
72 |
-
structured_llm = llm.with_structured_output(Insights)
|
73 |
-
prompt = ChatPromptTemplate.from_messages([system_message_template, UserMessage])
|
74 |
-
|
75 |
-
chain = prompt | structured_llm
|
76 |
-
|
77 |
-
new_insights = []
|
78 |
-
insights_data = []
|
79 |
-
|
80 |
-
while True:
|
81 |
-
# Invoke the LLM with the current chunk and existing insights
|
82 |
-
counter = 5 - len(new_insights)
|
83 |
-
new_insight_response = chain.invoke({"chunk": chunk, "existing_insights": new_insights, "counter": counter, "GSKGlossary": GSKGlossary, "topic":topic})
|
84 |
-
classification = selectClass(new_insight_response.insight)
|
85 |
-
# Append the new insight to the list
|
86 |
-
new_insights.append(new_insight_response.insight)
|
87 |
-
insights_data.append({
|
88 |
-
|
89 |
-
# "completed": new_insight_response.completed,
|
90 |
-
"classification": classification,
|
91 |
-
"insight": new_insight_response.insight,
|
92 |
-
"chunk": chunk
|
93 |
-
})
|
94 |
-
|
95 |
-
|
96 |
-
# Check if "completed" is True or the list of "new_insights" is >= 3
|
97 |
-
if new_insight_response.completed and len(new_insights) >= 3:
|
98 |
-
return pd.DataFrame(insights_data)
|
99 |
-
|
100 |
-
# If the list of "new_insights" reaches 5, return the list
|
101 |
-
if len(new_insights) == 5:
|
102 |
-
return pd.DataFrame(insights_data)
|
103 |
-
|
104 |
-
def selectClass(insight):
|
105 |
-
|
106 |
-
classification_system_message = read_file("classification_system_message")
|
107 |
-
classification_user_message = read_file("classification_user_message")
|
108 |
-
|
109 |
-
class InsightClassification(Enum):
|
110 |
-
IMPACT = "impact"
|
111 |
-
CONSULTATION = "consultation"
|
112 |
-
AWARENESS = "awareness"
|
113 |
-
|
114 |
-
llm = AzureChatOpenAI(
|
115 |
-
openai_api_version=OPENAI_API_VERSION,
|
116 |
-
openai_api_key=OPENAI_API_KEY,
|
117 |
-
azure_endpoint=OPENAI_API_BASE,
|
118 |
-
openai_api_type=OPENAI_API_TYPE,
|
119 |
-
deployment_name=OPENAI_MODEL,
|
120 |
-
temperature=0,
|
121 |
-
)
|
122 |
-
parser = EnumOutputParser(enum=InsightClassification)
|
123 |
-
system_message_template = SystemMessagePromptTemplate.from_template(classification_system_message)
|
124 |
-
|
125 |
-
# structured_llm = llm.with_structured_output(Insights)
|
126 |
-
prompt = ChatPromptTemplate.from_messages([system_message_template, classification_user_message]).partial(options=parser.get_format_instructions())
|
127 |
-
|
128 |
-
chain = prompt | llm | parser
|
129 |
-
|
130 |
-
result = chain.invoke({"insight": insight})
|
131 |
-
return result.value
|
132 |
-
|
133 |
-
def process_chunks(chunk, topic,source):
|
134 |
-
"""
|
135 |
-
Processes chunks from a specific dataframe column, invokes the get_structured function for each chunk,
|
136 |
-
and combines the resulting dataframes into one dataframe.
|
137 |
-
:param df: The dataframe containing chunks.
|
138 |
-
:param temp: Temperature parameter for the LLM.
|
139 |
-
:param SystemMessage: System message template.
|
140 |
-
:param UserMessage: User message template.
|
141 |
-
:param completedMessage: Completion message description.
|
142 |
-
:param insightMessage: Insight message description.
|
143 |
-
:param chunk_column: The name of the column containing text chunks to process.
|
144 |
-
:return: A combined dataframe of insights from all chunks.
|
145 |
-
"""
|
146 |
-
all_insights = []
|
147 |
-
|
148 |
-
for chunk in chunk["ChunkText"]:
|
149 |
-
insights_df = process_insight(chunk, topic,source)
|
150 |
-
all_insights.append(insights_df)
|
151 |
-
|
152 |
-
return pd.concat(all_insights, ignore_index=True)
|
153 |
-
|
154 |
-
|
155 |
-
def evaluation_llm(chunk, topic , source):
|
156 |
-
|
157 |
-
GSKGlossary = read_file("GSKGlossary")
|
158 |
-
if source == "intl":
|
159 |
-
SystemMessage = read_file("intl_eval_system_message")
|
160 |
-
UserMessage = read_file("intl_eval_user_message")
|
161 |
-
else:
|
162 |
-
SystemMessage = read_file("ext_eval_system_message")
|
163 |
-
UserMessage = read_file("ext_eval_user_message")
|
164 |
-
|
165 |
-
class Evaluate(BaseModel):
|
166 |
-
decision: bool = Field(description="True: The content of the document relates to the topic.False: The content of the document does not relate to the topic.")
|
167 |
-
justification: str = Field(description="Please justify your decision in a logical and structured way.")
|
168 |
-
|
169 |
-
llm = AzureChatOpenAI(
|
170 |
-
openai_api_version=OPENAI_API_VERSION,
|
171 |
-
openai_api_key=OPENAI_API_KEY,
|
172 |
-
azure_endpoint=OPENAI_API_BASE,
|
173 |
-
openai_api_type=OPENAI_API_TYPE,
|
174 |
-
deployment_name=OPENAI_MODEL,
|
175 |
-
temperature=0,
|
176 |
-
)
|
177 |
-
|
178 |
-
|
179 |
-
system_message_template = SystemMessagePromptTemplate.from_template(SystemMessage)
|
180 |
-
structured_llm = llm.with_structured_output(Evaluate)
|
181 |
-
|
182 |
-
# Create a chat prompt template combining system and human messages
|
183 |
-
prompt = ChatPromptTemplate.from_messages([system_message_template, UserMessage])
|
184 |
-
|
185 |
-
chain = prompt | structured_llm
|
186 |
-
|
187 |
-
return chain.invoke({
|
188 |
-
"chunk": chunk,
|
189 |
-
"topic": topic,
|
190 |
-
"GSKGlossary": GSKGlossary
|
191 |
-
})
|
192 |
-
|
193 |
-
def evaluation_process(df_chunks, topic,source):
|
194 |
-
"""
|
195 |
-
Iterates over chunks in the DataFrame and processes them using `get_structured`.
|
196 |
-
|
197 |
-
:param df_chunks: DataFrame containing chunks.
|
198 |
-
:param systemMessage: System message for evaluation.
|
199 |
-
:param userMessage: User message template for evaluation.
|
200 |
-
:param temp: Temperature setting for the model.
|
201 |
-
:param decisionMessage: Description for decision field.
|
202 |
-
:param justificationMessage: Description for justification field.
|
203 |
-
:return: Updated DataFrame with decision and justification columns and consensus value.
|
204 |
-
"""
|
205 |
-
decisions = []
|
206 |
-
justifications = []
|
207 |
-
|
208 |
-
# Avoid re-inserting columns if they already exist
|
209 |
-
if "Decision" in df_chunks.columns:
|
210 |
-
df_chunks = df_chunks.drop(columns=["Decision", "Justification"])
|
211 |
-
|
212 |
-
for _, chunk in df_chunks.iterrows():
|
213 |
-
result = evaluation_llm(chunk['ChunkText'], topic,source)
|
214 |
-
decisions.append("True" if result.decision else "False") # Convert bool to string
|
215 |
-
justifications.append(result.justification)
|
216 |
-
|
217 |
-
# Add new columns to the DataFrame
|
218 |
-
df_chunks.insert(0, "Decision", decisions)
|
219 |
-
df_chunks.insert(1, "Justification", justifications)
|
220 |
-
|
221 |
-
# Count all True/False values for consensus and get most frequent value
|
222 |
-
consensus_count = df_chunks["Decision"].value_counts()
|
223 |
-
consensus_value = consensus_count.idxmax() # Most frequently occurring value
|
224 |
-
|
225 |
-
return df_chunks, consensus_value, consensus_count
|
226 |
-
|
227 |
-
|
228 |
-
def process_compare(insight_df, sopChunk_df, topic):
|
229 |
-
|
230 |
-
GSKGlossary = read_file("GSKGlossary")
|
231 |
-
|
232 |
-
SystemMessage = read_file("compare_system_message")
|
233 |
-
UserMessage = read_file("compare_user_message")
|
234 |
-
|
235 |
-
# Define the structured output model
|
236 |
-
class Compare(BaseModel):
|
237 |
-
review: bool = Field(description="This field is used to indicate whether a review is needed")
|
238 |
-
justification: str = Field(description="This field is used to justify why a review is needed")
|
239 |
-
|
240 |
-
# Initialize the LLM
|
241 |
-
llm = AzureChatOpenAI(
|
242 |
-
openai_api_version=OPENAI_API_VERSION,
|
243 |
-
openai_api_key=OPENAI_API_KEY,
|
244 |
-
azure_endpoint=OPENAI_API_BASE,
|
245 |
-
openai_api_type=OPENAI_API_TYPE,
|
246 |
-
deployment_name=OPENAI_MODEL,
|
247 |
-
temperature=0,
|
248 |
-
)
|
249 |
-
|
250 |
-
# Create the structured output and prompt chain
|
251 |
-
system_message_template = SystemMessagePromptTemplate.from_template(SystemMessage)
|
252 |
-
structured_llm = llm.with_structured_output(Compare)
|
253 |
-
prompt = ChatPromptTemplate.from_messages([system_message_template, UserMessage])
|
254 |
-
chain = prompt | structured_llm
|
255 |
-
|
256 |
-
compare_data = []
|
257 |
-
|
258 |
-
# Iterate over sopChunk_df and insight_df to process "ChunkText" and "insight"
|
259 |
-
for sopChunk_index, sopChunk_row in sopChunk_df.iterrows():
|
260 |
-
sop_chunk_text = sopChunk_row["ChunkText"] # Extract the ChunkText column
|
261 |
-
for insight_index, insight_row in insight_df.iterrows():
|
262 |
-
insight_text = insight_row["insight"] # Extract the insight column
|
263 |
-
|
264 |
-
# Invoke the LLM with the extracted data
|
265 |
-
compare_response = chain.invoke({
|
266 |
-
"sopChunk": sop_chunk_text,
|
267 |
-
"insight": insight_text,
|
268 |
-
"topic": topic,
|
269 |
-
"GSKGlossary": GSKGlossary
|
270 |
-
})
|
271 |
-
|
272 |
-
# Append the response to insights_data
|
273 |
-
compare_data.append({
|
274 |
-
"ReviewNeeded": compare_response.review,
|
275 |
-
"Justification": compare_response.justification,
|
276 |
-
"SOP": sop_chunk_text,
|
277 |
-
"Insight": insight_text
|
278 |
-
})
|
279 |
-
|
280 |
-
# Return the insights as a single DataFrame
|
281 |
-
print(compare_data)
|
282 |
-
return pd.DataFrame(compare_data)
|
283 |
-
|
284 |
-
def risk_score_process(compare_df, topic):
|
285 |
-
|
286 |
-
GSKGlossary = read_file("GSKGlossary")
|
287 |
-
SystemMessage = read_file("risk_scoring_system_message")
|
288 |
-
UserMessage = read_file("risk_scoring_user_message")
|
289 |
-
|
290 |
-
# Define the Enum for predefined options
|
291 |
-
class RiskClassification(str, Enum):
|
292 |
-
HIGH = "high"
|
293 |
-
MEDIUM = "medium"
|
294 |
-
LOW = "low"
|
295 |
-
|
296 |
-
# Define the Pydantic model for the structured output
|
297 |
-
class Risk(BaseModel):
|
298 |
-
risk_level: RiskClassification = Field(
|
299 |
-
description="The selected classification option."
|
300 |
-
)
|
301 |
-
justification: str = Field(
|
302 |
-
description="Justify the reason for choosing this risk classification."
|
303 |
-
)
|
304 |
-
advice: str = Field(
|
305 |
-
description="Suggestions for changes that could be made to the standard operating procedure to mitigat the risk."
|
306 |
-
)
|
307 |
-
|
308 |
-
|
309 |
-
llm = AzureChatOpenAI(
|
310 |
-
openai_api_version=OPENAI_API_VERSION,
|
311 |
-
openai_api_key=OPENAI_API_KEY,
|
312 |
-
azure_endpoint=OPENAI_API_BASE,
|
313 |
-
openai_api_type=OPENAI_API_TYPE,
|
314 |
-
deployment_name=OPENAI_MODEL,
|
315 |
-
temperature=0,
|
316 |
-
)
|
317 |
-
|
318 |
-
system_message_template = SystemMessagePromptTemplate.from_template(SystemMessage)
|
319 |
-
structured_llm = llm.with_structured_output(Risk)
|
320 |
-
prompt = ChatPromptTemplate.from_messages([system_message_template, UserMessage])
|
321 |
-
|
322 |
-
chain = prompt | structured_llm
|
323 |
-
|
324 |
-
risk_data = []
|
325 |
-
|
326 |
-
|
327 |
-
# Iterate over sopChunk_df and insight_df to process "ChunkText" and "insight"
|
328 |
-
for index, row in compare_df.iterrows():
|
329 |
-
|
330 |
-
# Invoke the LLM with the extracted data
|
331 |
-
risk_response = chain.invoke({
|
332 |
-
"comparison": row['Justification'],
|
333 |
-
"insight": row['Insight'],
|
334 |
-
"SOPchunk":row['SOP'],
|
335 |
-
"topic": topic
|
336 |
-
})
|
337 |
-
|
338 |
-
# Append the response to insights_data
|
339 |
-
risk_data.append({
|
340 |
-
"RiskLevel": risk_response.risk_level,
|
341 |
-
"Justification": risk_response.justification,
|
342 |
-
"advice": risk_response.advice,
|
343 |
-
"comparison": row['Justification'],
|
344 |
-
"insight": row['Insight'],
|
345 |
-
"SOPchunk":row['SOP']
|
346 |
-
})
|
347 |
-
|
348 |
-
# Return the insights as a single DataFrame
|
349 |
return pd.DataFrame(risk_data)
|
|
|
1 |
+
import streamlit as st
|
2 |
+
import os
|
3 |
+
import pandas as pd
|
4 |
+
# from langchain.chat_models import AzureChatOpenAI
|
5 |
+
from langchain_openai import AzureChatOpenAI
|
6 |
+
from langchain_core.output_parsers import StrOutputParser, PydanticOutputParser
|
7 |
+
from langchain_core.prompts.chat import SystemMessagePromptTemplate, HumanMessagePromptTemplate, ChatPromptTemplate
|
8 |
+
from pydantic import BaseModel, Field, validator
|
9 |
+
from langchain.output_parsers.enum import EnumOutputParser
|
10 |
+
from langchain_core.prompts import PromptTemplate
|
11 |
+
from enum import Enum
|
12 |
+
|
13 |
+
|
14 |
+
#os.environ["LANGCHAIN_TRACING_V2"]="true"
|
15 |
+
#os.environ["LANGCHAIN_ENDPOINT"]="https://api.smith.langchain.com"
|
16 |
+
#LANGCHAIN_API_KEY = st.secrets['LANGCHAIN_API_KEY']
|
17 |
+
#os.environ["LANGCHAIN_PROJECT"]="UC2e2e"
|
18 |
+
|
19 |
+
# LLM Langchain Definition
|
20 |
+
OPENAI_API_KEY = st.secrets['OPENAI_API_KEY']
|
21 |
+
OPENAI_API_TYPE = "azure"
|
22 |
+
OPENAI_API_BASE = "https://davidfearn-gpt4.openai.azure.com"
|
23 |
+
OPENAI_API_VERSION = "2024-08-01-preview"
|
24 |
+
OPENAI_MODEL = "gpt-4o-mini"
|
25 |
+
|
26 |
+
|
27 |
+
# Function to read file contents
|
28 |
+
def read_file(file):
|
29 |
+
"""
|
30 |
+
Reads the content of a text file and returns it as a string.
|
31 |
+
:param file: The file name to read from the 'assets' directory.
|
32 |
+
:return: The content of the file as a string or None if an error occurs.
|
33 |
+
"""
|
34 |
+
fp = f"assets/{file}.md"
|
35 |
+
try:
|
36 |
+
with open(fp, 'r', encoding='utf-8') as file:
|
37 |
+
content = file.read()
|
38 |
+
return content
|
39 |
+
except FileNotFoundError:
|
40 |
+
print(f"The file at {fp} was not found.")
|
41 |
+
except IOError:
|
42 |
+
print(f"An error occurred while reading the file at {fp}.")
|
43 |
+
return None
|
44 |
+
|
45 |
+
# Function to generate structured insights
|
46 |
+
def process_insight(chunk, topic,source):
|
47 |
+
|
48 |
+
GSKGlossary = read_file("GSKGlossary")
|
49 |
+
if source== "intl":
|
50 |
+
SystemMessage = read_file("intl_insight_system_message")
|
51 |
+
UserMessage = read_file("intl_insight_user_message")
|
52 |
+
else:
|
53 |
+
SystemMessage = read_file("ext_insight_system_message")
|
54 |
+
UserMessage = read_file("ext_insight_user_message")
|
55 |
+
|
56 |
+
|
57 |
+
class Insights(BaseModel):
|
58 |
+
completed: bool = Field(description="This field is used to indicate that you think the number of insights has been completed")
|
59 |
+
insight: str = Field(description="This field is used to return the MECE insight in string format")
|
60 |
+
|
61 |
+
|
62 |
+
llm = AzureChatOpenAI(
|
63 |
+
openai_api_version=OPENAI_API_VERSION,
|
64 |
+
openai_api_key=OPENAI_API_KEY,
|
65 |
+
azure_endpoint=OPENAI_API_BASE,
|
66 |
+
openai_api_type=OPENAI_API_TYPE,
|
67 |
+
deployment_name=OPENAI_MODEL,
|
68 |
+
temperature=0,
|
69 |
+
)
|
70 |
+
|
71 |
+
system_message_template = SystemMessagePromptTemplate.from_template(SystemMessage)
|
72 |
+
structured_llm = llm.with_structured_output(Insights)
|
73 |
+
prompt = ChatPromptTemplate.from_messages([system_message_template, UserMessage])
|
74 |
+
|
75 |
+
chain = prompt | structured_llm
|
76 |
+
|
77 |
+
new_insights = []
|
78 |
+
insights_data = []
|
79 |
+
|
80 |
+
while True:
|
81 |
+
# Invoke the LLM with the current chunk and existing insights
|
82 |
+
counter = 5 - len(new_insights)
|
83 |
+
new_insight_response = chain.invoke({"chunk": chunk, "existing_insights": new_insights, "counter": counter, "GSKGlossary": GSKGlossary, "topic":topic})
|
84 |
+
classification = selectClass(new_insight_response.insight)
|
85 |
+
# Append the new insight to the list
|
86 |
+
new_insights.append(new_insight_response.insight)
|
87 |
+
insights_data.append({
|
88 |
+
|
89 |
+
# "completed": new_insight_response.completed,
|
90 |
+
"classification": classification,
|
91 |
+
"insight": new_insight_response.insight,
|
92 |
+
"chunk": chunk
|
93 |
+
})
|
94 |
+
|
95 |
+
|
96 |
+
# Check if "completed" is True or the list of "new_insights" is >= 3
|
97 |
+
if new_insight_response.completed and len(new_insights) >= 3:
|
98 |
+
return pd.DataFrame(insights_data)
|
99 |
+
|
100 |
+
# If the list of "new_insights" reaches 5, return the list
|
101 |
+
if len(new_insights) == 5:
|
102 |
+
return pd.DataFrame(insights_data)
|
103 |
+
|
104 |
+
def selectClass(insight):
|
105 |
+
|
106 |
+
classification_system_message = read_file("classification_system_message")
|
107 |
+
classification_user_message = read_file("classification_user_message")
|
108 |
+
|
109 |
+
class InsightClassification(Enum):
|
110 |
+
IMPACT = "impact"
|
111 |
+
CONSULTATION = "consultation"
|
112 |
+
AWARENESS = "awareness"
|
113 |
+
|
114 |
+
llm = AzureChatOpenAI(
|
115 |
+
openai_api_version=OPENAI_API_VERSION,
|
116 |
+
openai_api_key=OPENAI_API_KEY,
|
117 |
+
azure_endpoint=OPENAI_API_BASE,
|
118 |
+
openai_api_type=OPENAI_API_TYPE,
|
119 |
+
deployment_name=OPENAI_MODEL,
|
120 |
+
temperature=0,
|
121 |
+
)
|
122 |
+
parser = EnumOutputParser(enum=InsightClassification)
|
123 |
+
system_message_template = SystemMessagePromptTemplate.from_template(classification_system_message)
|
124 |
+
|
125 |
+
# structured_llm = llm.with_structured_output(Insights)
|
126 |
+
prompt = ChatPromptTemplate.from_messages([system_message_template, classification_user_message]).partial(options=parser.get_format_instructions())
|
127 |
+
|
128 |
+
chain = prompt | llm | parser
|
129 |
+
|
130 |
+
result = chain.invoke({"insight": insight})
|
131 |
+
return result.value
|
132 |
+
|
133 |
+
def process_chunks(chunk, topic,source):
|
134 |
+
"""
|
135 |
+
Processes chunks from a specific dataframe column, invokes the get_structured function for each chunk,
|
136 |
+
and combines the resulting dataframes into one dataframe.
|
137 |
+
:param df: The dataframe containing chunks.
|
138 |
+
:param temp: Temperature parameter for the LLM.
|
139 |
+
:param SystemMessage: System message template.
|
140 |
+
:param UserMessage: User message template.
|
141 |
+
:param completedMessage: Completion message description.
|
142 |
+
:param insightMessage: Insight message description.
|
143 |
+
:param chunk_column: The name of the column containing text chunks to process.
|
144 |
+
:return: A combined dataframe of insights from all chunks.
|
145 |
+
"""
|
146 |
+
all_insights = []
|
147 |
+
|
148 |
+
for chunk in chunk["ChunkText"]:
|
149 |
+
insights_df = process_insight(chunk, topic,source)
|
150 |
+
all_insights.append(insights_df)
|
151 |
+
|
152 |
+
return pd.concat(all_insights, ignore_index=True)
|
153 |
+
|
154 |
+
|
155 |
+
def evaluation_llm(chunk, topic , source):
|
156 |
+
|
157 |
+
GSKGlossary = read_file("GSKGlossary")
|
158 |
+
if source == "intl":
|
159 |
+
SystemMessage = read_file("intl_eval_system_message")
|
160 |
+
UserMessage = read_file("intl_eval_user_message")
|
161 |
+
else:
|
162 |
+
SystemMessage = read_file("ext_eval_system_message")
|
163 |
+
UserMessage = read_file("ext_eval_user_message")
|
164 |
+
|
165 |
+
class Evaluate(BaseModel):
|
166 |
+
decision: bool = Field(description="True: The content of the document relates to the topic.False: The content of the document does not relate to the topic.")
|
167 |
+
justification: str = Field(description="Please justify your decision in a logical and structured way.")
|
168 |
+
|
169 |
+
llm = AzureChatOpenAI(
|
170 |
+
openai_api_version=OPENAI_API_VERSION,
|
171 |
+
openai_api_key=OPENAI_API_KEY,
|
172 |
+
azure_endpoint=OPENAI_API_BASE,
|
173 |
+
openai_api_type=OPENAI_API_TYPE,
|
174 |
+
deployment_name=OPENAI_MODEL,
|
175 |
+
temperature=0,
|
176 |
+
)
|
177 |
+
|
178 |
+
|
179 |
+
system_message_template = SystemMessagePromptTemplate.from_template(SystemMessage)
|
180 |
+
structured_llm = llm.with_structured_output(Evaluate)
|
181 |
+
|
182 |
+
# Create a chat prompt template combining system and human messages
|
183 |
+
prompt = ChatPromptTemplate.from_messages([system_message_template, UserMessage])
|
184 |
+
|
185 |
+
chain = prompt | structured_llm
|
186 |
+
|
187 |
+
return chain.invoke({
|
188 |
+
"chunk": chunk,
|
189 |
+
"topic": topic,
|
190 |
+
"GSKGlossary": GSKGlossary
|
191 |
+
})
|
192 |
+
|
193 |
+
def evaluation_process(df_chunks, topic,source):
|
194 |
+
"""
|
195 |
+
Iterates over chunks in the DataFrame and processes them using `get_structured`.
|
196 |
+
|
197 |
+
:param df_chunks: DataFrame containing chunks.
|
198 |
+
:param systemMessage: System message for evaluation.
|
199 |
+
:param userMessage: User message template for evaluation.
|
200 |
+
:param temp: Temperature setting for the model.
|
201 |
+
:param decisionMessage: Description for decision field.
|
202 |
+
:param justificationMessage: Description for justification field.
|
203 |
+
:return: Updated DataFrame with decision and justification columns and consensus value.
|
204 |
+
"""
|
205 |
+
decisions = []
|
206 |
+
justifications = []
|
207 |
+
|
208 |
+
# Avoid re-inserting columns if they already exist
|
209 |
+
if "Decision" in df_chunks.columns:
|
210 |
+
df_chunks = df_chunks.drop(columns=["Decision", "Justification"])
|
211 |
+
|
212 |
+
for _, chunk in df_chunks.iterrows():
|
213 |
+
result = evaluation_llm(chunk['ChunkText'], topic,source)
|
214 |
+
decisions.append("True" if result.decision else "False") # Convert bool to string
|
215 |
+
justifications.append(result.justification)
|
216 |
+
|
217 |
+
# Add new columns to the DataFrame
|
218 |
+
df_chunks.insert(0, "Decision", decisions)
|
219 |
+
df_chunks.insert(1, "Justification", justifications)
|
220 |
+
|
221 |
+
# Count all True/False values for consensus and get most frequent value
|
222 |
+
consensus_count = df_chunks["Decision"].value_counts()
|
223 |
+
consensus_value = consensus_count.idxmax() # Most frequently occurring value
|
224 |
+
|
225 |
+
return df_chunks, consensus_value, consensus_count
|
226 |
+
|
227 |
+
|
228 |
+
def process_compare(insight_df, sopChunk_df, topic):
|
229 |
+
|
230 |
+
GSKGlossary = read_file("GSKGlossary")
|
231 |
+
|
232 |
+
SystemMessage = read_file("compare_system_message")
|
233 |
+
UserMessage = read_file("compare_user_message")
|
234 |
+
|
235 |
+
# Define the structured output model
|
236 |
+
class Compare(BaseModel):
|
237 |
+
review: bool = Field(description="This field is used to indicate whether a review is needed")
|
238 |
+
justification: str = Field(description="This field is used to justify why a review is needed")
|
239 |
+
|
240 |
+
# Initialize the LLM
|
241 |
+
llm = AzureChatOpenAI(
|
242 |
+
openai_api_version=OPENAI_API_VERSION,
|
243 |
+
openai_api_key=OPENAI_API_KEY,
|
244 |
+
azure_endpoint=OPENAI_API_BASE,
|
245 |
+
openai_api_type=OPENAI_API_TYPE,
|
246 |
+
deployment_name=OPENAI_MODEL,
|
247 |
+
temperature=0,
|
248 |
+
)
|
249 |
+
|
250 |
+
# Create the structured output and prompt chain
|
251 |
+
system_message_template = SystemMessagePromptTemplate.from_template(SystemMessage)
|
252 |
+
structured_llm = llm.with_structured_output(Compare)
|
253 |
+
prompt = ChatPromptTemplate.from_messages([system_message_template, UserMessage])
|
254 |
+
chain = prompt | structured_llm
|
255 |
+
|
256 |
+
compare_data = []
|
257 |
+
|
258 |
+
# Iterate over sopChunk_df and insight_df to process "ChunkText" and "insight"
|
259 |
+
for sopChunk_index, sopChunk_row in sopChunk_df.iterrows():
|
260 |
+
sop_chunk_text = sopChunk_row["ChunkText"] # Extract the ChunkText column
|
261 |
+
for insight_index, insight_row in insight_df.iterrows():
|
262 |
+
insight_text = insight_row["insight"] # Extract the insight column
|
263 |
+
|
264 |
+
# Invoke the LLM with the extracted data
|
265 |
+
compare_response = chain.invoke({
|
266 |
+
"sopChunk": sop_chunk_text,
|
267 |
+
"insight": insight_text,
|
268 |
+
"topic": topic,
|
269 |
+
"GSKGlossary": GSKGlossary
|
270 |
+
})
|
271 |
+
|
272 |
+
# Append the response to insights_data
|
273 |
+
compare_data.append({
|
274 |
+
"ReviewNeeded": compare_response.review,
|
275 |
+
"Justification": compare_response.justification,
|
276 |
+
"SOP": sop_chunk_text,
|
277 |
+
"Insight": insight_text
|
278 |
+
})
|
279 |
+
|
280 |
+
# Return the insights as a single DataFrame
|
281 |
+
print(compare_data)
|
282 |
+
return pd.DataFrame(compare_data)
|
283 |
+
|
284 |
+
def risk_score_process(compare_df, topic):
|
285 |
+
|
286 |
+
GSKGlossary = read_file("GSKGlossary")
|
287 |
+
SystemMessage = read_file("risk_scoring_system_message")
|
288 |
+
UserMessage = read_file("risk_scoring_user_message")
|
289 |
+
|
290 |
+
# Define the Enum for predefined options
|
291 |
+
class RiskClassification(str, Enum):
|
292 |
+
HIGH = "high"
|
293 |
+
MEDIUM = "medium"
|
294 |
+
LOW = "low"
|
295 |
+
|
296 |
+
# Define the Pydantic model for the structured output
|
297 |
+
class Risk(BaseModel):
|
298 |
+
risk_level: RiskClassification = Field(
|
299 |
+
description="The selected classification option."
|
300 |
+
)
|
301 |
+
justification: str = Field(
|
302 |
+
description="Justify the reason for choosing this risk classification."
|
303 |
+
)
|
304 |
+
advice: str = Field(
|
305 |
+
description="Suggestions for changes that could be made to the standard operating procedure to mitigat the risk."
|
306 |
+
)
|
307 |
+
|
308 |
+
|
309 |
+
llm = AzureChatOpenAI(
|
310 |
+
openai_api_version=OPENAI_API_VERSION,
|
311 |
+
openai_api_key=OPENAI_API_KEY,
|
312 |
+
azure_endpoint=OPENAI_API_BASE,
|
313 |
+
openai_api_type=OPENAI_API_TYPE,
|
314 |
+
deployment_name=OPENAI_MODEL,
|
315 |
+
temperature=0,
|
316 |
+
)
|
317 |
+
|
318 |
+
system_message_template = SystemMessagePromptTemplate.from_template(SystemMessage)
|
319 |
+
structured_llm = llm.with_structured_output(Risk)
|
320 |
+
prompt = ChatPromptTemplate.from_messages([system_message_template, UserMessage])
|
321 |
+
|
322 |
+
chain = prompt | structured_llm
|
323 |
+
|
324 |
+
risk_data = []
|
325 |
+
|
326 |
+
|
327 |
+
# Iterate over sopChunk_df and insight_df to process "ChunkText" and "insight"
|
328 |
+
for index, row in compare_df.iterrows():
|
329 |
+
|
330 |
+
# Invoke the LLM with the extracted data
|
331 |
+
risk_response = chain.invoke({
|
332 |
+
"comparison": row['Justification'],
|
333 |
+
"insight": row['Insight'],
|
334 |
+
"SOPchunk":row['SOP'],
|
335 |
+
"topic": topic
|
336 |
+
})
|
337 |
+
|
338 |
+
# Append the response to insights_data
|
339 |
+
risk_data.append({
|
340 |
+
"RiskLevel": risk_response.risk_level,
|
341 |
+
"Justification": risk_response.justification,
|
342 |
+
"advice": risk_response.advice,
|
343 |
+
"comparison": row['Justification'],
|
344 |
+
"insight": row['Insight'],
|
345 |
+
"SOPchunk":row['SOP']
|
346 |
+
})
|
347 |
+
|
348 |
+
# Return the insights as a single DataFrame
|
349 |
return pd.DataFrame(risk_data)
|