Spaces:
Sleeping
Sleeping
TillLangbein
commited on
Commit
·
c3057bf
1
Parent(s):
80e25da
Add application file
Browse files- .cache.db +0 -0
- __pycache__/app.cpython-312.pyc +0 -0
- __pycache__/prompts.cpython-312.pyc +0 -0
- app.py +494 -0
- logo.png +0 -0
- prompts.py +69 -0
- requirements.txt +7 -0
- style.css +65 -0
.cache.db
ADDED
Binary file (332 kB). View file
|
|
__pycache__/app.cpython-312.pyc
ADDED
Binary file (173 Bytes). View file
|
|
__pycache__/prompts.cpython-312.pyc
ADDED
Binary file (3.56 kB). View file
|
|
app.py
ADDED
@@ -0,0 +1,494 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import getpass
|
2 |
+
import os
|
3 |
+
import random
|
4 |
+
|
5 |
+
from langchain_openai import ChatOpenAI
|
6 |
+
from langchain_core.globals import set_llm_cache
|
7 |
+
from langchain_community.cache import SQLiteCache
|
8 |
+
from langchain_community.vectorstores import FAISS
|
9 |
+
from langchain_openai import OpenAIEmbeddings
|
10 |
+
from langgraph.graph import END, StateGraph, START
|
11 |
+
from langchain_core.output_parsers import StrOutputParser
|
12 |
+
|
13 |
+
from typing import List
|
14 |
+
from typing_extensions import TypedDict
|
15 |
+
import gradio as gr
|
16 |
+
from pydantic import BaseModel, Field
|
17 |
+
|
18 |
+
from prompts import IMPROVE_PROMPT, RELEVANCE_PROMPT, ANSWER_PROMPT, HALLUCINATION_PROMPT, RESOLVER_PROMPT, REWRITER_PROMPT
|
19 |
+
|
20 |
+
TOPICS = [
|
21 |
+
"ICT strategy management",
|
22 |
+
"IT governance management & internal controls system",
|
23 |
+
"Internal audit & compliance management",
|
24 |
+
"ICT asset & architecture management",
|
25 |
+
"ICT risk management",
|
26 |
+
"Information security & human resource security management",
|
27 |
+
"IT configuration management",
|
28 |
+
"Cryptography, certificates & key management",
|
29 |
+
"Secure network & infrastructure management",
|
30 |
+
"Backup",
|
31 |
+
"Security testing",
|
32 |
+
"Threat-led penetration testing",
|
33 |
+
"Logging",
|
34 |
+
"Data and ICT system security",
|
35 |
+
"Physical and environmental security",
|
36 |
+
"Vulnerability & patch management",
|
37 |
+
"Identity and access management",
|
38 |
+
"ICT change management",
|
39 |
+
"IT project & project portfolio management",
|
40 |
+
"Acquisition, development & maintenance of ICT systems & EUA",
|
41 |
+
"ICT incident management",
|
42 |
+
"Monitoring, availability, capacity & performance management",
|
43 |
+
"ICT outsourcing & third-party risk management",
|
44 |
+
"Subcontracting management",
|
45 |
+
"ICT provider & service level management",
|
46 |
+
"ICT business continuity management"
|
47 |
+
]
|
48 |
+
|
49 |
+
class GradeDocuments(BaseModel):
|
50 |
+
"""Binary score for relevance check on retrieved documents."""
|
51 |
+
|
52 |
+
binary_score: str = Field(
|
53 |
+
description="Documents are relevant to the question, 'yes' or 'no'"
|
54 |
+
)
|
55 |
+
|
56 |
+
class GradeHallucinations(BaseModel):
|
57 |
+
"""Binary score for hallucination present in generation answer."""
|
58 |
+
|
59 |
+
binary_score: str = Field(
|
60 |
+
description="Answer is grounded in the facts, 'yes' or 'no'"
|
61 |
+
)
|
62 |
+
|
63 |
+
class GradeAnswer(BaseModel):
|
64 |
+
"""Binary score to assess answer addresses question."""
|
65 |
+
|
66 |
+
binary_score: str = Field(
|
67 |
+
description="Answer addresses the question, 'yes' or 'no'"
|
68 |
+
)
|
69 |
+
|
70 |
+
class GraphState(TypedDict):
|
71 |
+
"""
|
72 |
+
Represents the state of our graph.
|
73 |
+
|
74 |
+
Attributes:
|
75 |
+
question: question
|
76 |
+
generation: LLM generation
|
77 |
+
documents: list of documents
|
78 |
+
"""
|
79 |
+
|
80 |
+
question: str
|
81 |
+
selected_sources: List[List[bool]]
|
82 |
+
generation: str
|
83 |
+
documents: List[str]
|
84 |
+
fitting_documents: List[str]
|
85 |
+
dora_docs: List[str]
|
86 |
+
dora_rts_docs: List[str]
|
87 |
+
dora_news_docs: List[str]
|
88 |
+
|
89 |
+
def _set_env(var: str):
|
90 |
+
if os.environ.get(var):
|
91 |
+
return
|
92 |
+
os.environ[var] = getpass.getpass(var + ":")
|
93 |
+
|
94 |
+
def load_vectorstores(paths: list):
|
95 |
+
# The dora vectorstore
|
96 |
+
embd = OpenAIEmbeddings()
|
97 |
+
|
98 |
+
vectorstores = [FAISS.load_local(path, embd, allow_dangerous_deserialization=True) for path in paths]
|
99 |
+
retrievers = [vectorstore.as_retriever(search_type="mmr", search_kwargs={
|
100 |
+
"k": 7,
|
101 |
+
"fetch_k": 10,
|
102 |
+
"score_threshold": 0.7,
|
103 |
+
}) for vectorstore in vectorstores]
|
104 |
+
|
105 |
+
return retrievers
|
106 |
+
|
107 |
+
# Put all chains in fuctions
|
108 |
+
def dora_rewrite(state):
|
109 |
+
"""
|
110 |
+
Rewrites the question to fit dora wording
|
111 |
+
|
112 |
+
Args:
|
113 |
+
state (dict): The current graph state
|
114 |
+
|
115 |
+
Returns:
|
116 |
+
state (dict): New key added to state, documents, that contains retrieved documents
|
117 |
+
"""
|
118 |
+
print("---TRANSLATE TO DORA---")
|
119 |
+
question = state["question"]
|
120 |
+
|
121 |
+
new_question = dora_question_rewriter.invoke({"question": question, "topics": TOPICS})
|
122 |
+
|
123 |
+
if new_question == "Thats an interesting question, but I dont think I can answer it based on my Dora knowledge.":
|
124 |
+
return {"question": new_question, "generation": new_question}
|
125 |
+
else:
|
126 |
+
return {"question": new_question}
|
127 |
+
|
128 |
+
def retrieve(state):
|
129 |
+
"""
|
130 |
+
Retrieve documents
|
131 |
+
|
132 |
+
Args:
|
133 |
+
state (dict): The current graph state
|
134 |
+
|
135 |
+
Returns:
|
136 |
+
state (dict): New key added to state, documents, that contains retrieved documents
|
137 |
+
"""
|
138 |
+
print("---RETRIEVE---")
|
139 |
+
question = state["question"]
|
140 |
+
selected_sources = state["selected_sources"]
|
141 |
+
|
142 |
+
# Retrieval
|
143 |
+
documents = []
|
144 |
+
if selected_sources[0]:
|
145 |
+
documents.extend(dora_retriever.invoke(question))
|
146 |
+
if selected_sources[1]:
|
147 |
+
documents.extend(dora_rts_retriever.invoke(question))
|
148 |
+
if selected_sources[2]:
|
149 |
+
documents.extend(dora_news_retriever.invoke(question))
|
150 |
+
|
151 |
+
return {"documents": documents, "question": question}
|
152 |
+
|
153 |
+
def grade_documents(state):
|
154 |
+
"""
|
155 |
+
Determines whether the retrieved documents are relevant to the question.
|
156 |
+
|
157 |
+
Args:
|
158 |
+
state (dict): The current graph state
|
159 |
+
|
160 |
+
Returns:
|
161 |
+
state (dict): Updates documents key with only filtered relevant documents
|
162 |
+
"""
|
163 |
+
|
164 |
+
print("---CHECK DOCUMENTS RELEVANCE TO QUESTION---")
|
165 |
+
question = state["question"]
|
166 |
+
documents = state["documents"]
|
167 |
+
fitting_documents = state["fitting_documents"] if "fitting_documents" in state else []
|
168 |
+
|
169 |
+
|
170 |
+
# Score each doc
|
171 |
+
for d in documents:
|
172 |
+
score = retrieval_grader.invoke(
|
173 |
+
{"question": question, "document": d.page_content}
|
174 |
+
)
|
175 |
+
grade = score.binary_score
|
176 |
+
if grade == "yes":
|
177 |
+
#print("---GRADE: DOCUMENT RELEVANT---")
|
178 |
+
if d in fitting_documents:
|
179 |
+
#print(f"---Document {d.page_content} already in fitting documents---")
|
180 |
+
continue
|
181 |
+
fitting_documents.append(d)
|
182 |
+
else:
|
183 |
+
#print("---GRADE: DOCUMENT NOT RELEVANT---")
|
184 |
+
continue
|
185 |
+
|
186 |
+
return {"fitting_documents": fitting_documents}
|
187 |
+
|
188 |
+
def generate(state):
|
189 |
+
"""
|
190 |
+
Generate answer
|
191 |
+
|
192 |
+
Args:
|
193 |
+
state (dict): The current graph state
|
194 |
+
|
195 |
+
Returns:
|
196 |
+
state (dict): New key added to state, generation, that contains LLM generation
|
197 |
+
"""
|
198 |
+
print("---GENERATE---")
|
199 |
+
question = state["question"]
|
200 |
+
fitting_documents = state["fitting_documents"]
|
201 |
+
|
202 |
+
dora_docs = [d for d in fitting_documents if d.metadata["source"].startswith("Dora")]
|
203 |
+
dora_rts_docs = [d for d in fitting_documents if d.metadata["source"].startswith("Commission")]
|
204 |
+
dora_news_docs = [d for d in fitting_documents if d.metadata["source"].startswith("https")]
|
205 |
+
|
206 |
+
# RAG generation
|
207 |
+
generation = answer_chain.invoke({"context": fitting_documents, "question": question})
|
208 |
+
return {"generation": generation, "dora_docs": dora_docs, "dora_rts_docs": dora_rts_docs, "dora_news_docs": dora_news_docs}
|
209 |
+
|
210 |
+
def transform_query(state):
|
211 |
+
"""
|
212 |
+
Transform the query to produce a better question.
|
213 |
+
|
214 |
+
Args:
|
215 |
+
state (dict): The current graph state
|
216 |
+
|
217 |
+
Returns:
|
218 |
+
state (dict): Updates question key with a re-phrased question
|
219 |
+
"""
|
220 |
+
|
221 |
+
print("---TRANSFORM QUERY---")
|
222 |
+
question = state["question"]
|
223 |
+
|
224 |
+
# Re-write question
|
225 |
+
better_question = question_rewriter.invoke({"question": question})
|
226 |
+
print(f"{better_question =}")
|
227 |
+
return {"question": better_question}
|
228 |
+
|
229 |
+
### Edges ###
|
230 |
+
def suitable_question(state):
|
231 |
+
"""
|
232 |
+
Determines whether the question is suitable.
|
233 |
+
|
234 |
+
Args:
|
235 |
+
state (dict): The current graph state
|
236 |
+
|
237 |
+
Returns:
|
238 |
+
str: Binary decision for next node to call
|
239 |
+
"""
|
240 |
+
|
241 |
+
print("---ASSESSING THE QUESTION---")
|
242 |
+
question = state["question"]
|
243 |
+
#print(f"{question = }")
|
244 |
+
if question == "Thats an interesting question, but I dont think I can answer it based on my Dora knowledge.":
|
245 |
+
return "end"
|
246 |
+
else:
|
247 |
+
return "retrieve"
|
248 |
+
|
249 |
+
def decide_to_generate(state):
|
250 |
+
"""
|
251 |
+
Determines whether to generate an answer, or re-generate a question.
|
252 |
+
|
253 |
+
Args:
|
254 |
+
state (dict): The current graph state
|
255 |
+
|
256 |
+
Returns:
|
257 |
+
str: Binary decision for next node to call
|
258 |
+
"""
|
259 |
+
|
260 |
+
print("---ASSESS GRADED DOCUMENTS---")
|
261 |
+
fitting_documents = state["fitting_documents"]
|
262 |
+
|
263 |
+
if not fitting_documents:
|
264 |
+
# All documents have been filtered check_relevance
|
265 |
+
# We will re-generate a new query
|
266 |
+
print(
|
267 |
+
"---DECISION: ALL DOCUMENTS ARE IRRELEVANT TO QUESTION, TRANSFORM QUERY---"
|
268 |
+
)
|
269 |
+
return "transform_query"
|
270 |
+
else:
|
271 |
+
# We have relevant documents, so generate answer
|
272 |
+
print(f"---DECISION: GENERATE WITH {len(fitting_documents)} DOCUMENTS---")
|
273 |
+
return "generate"
|
274 |
+
|
275 |
+
def grade_generation_v_documents_and_question(state):
|
276 |
+
"""
|
277 |
+
Determines whether the generation is grounded in the document and answers question.
|
278 |
+
|
279 |
+
Args:
|
280 |
+
state (dict): The current graph state
|
281 |
+
|
282 |
+
Returns:
|
283 |
+
str: Decision for next node to call
|
284 |
+
"""
|
285 |
+
|
286 |
+
print("---CHECK HALLUCINATIONS---")
|
287 |
+
question = state["question"]
|
288 |
+
fitting_documents = state["fitting_documents"]
|
289 |
+
generation = state["generation"]
|
290 |
+
|
291 |
+
score = hallucination_grader.invoke(
|
292 |
+
{"documents": fitting_documents, "generation": generation}
|
293 |
+
)
|
294 |
+
grade = score.binary_score
|
295 |
+
|
296 |
+
# Check hallucination
|
297 |
+
if grade == "yes":
|
298 |
+
print("---DECISION: GENERATION IS GROUNDED IN DOCUMENTS---")
|
299 |
+
# Check question-answering
|
300 |
+
print("---GRADE GENERATION vs QUESTION---")
|
301 |
+
score = answer_grader.invoke({"question": question, "generation": generation})
|
302 |
+
grade = score.binary_score
|
303 |
+
if grade == "yes":
|
304 |
+
print("---DECISION: GENERATION ADDRESSES QUESTION---")
|
305 |
+
return "useful"
|
306 |
+
else:
|
307 |
+
print("---DECISION: GENERATION DOES NOT ADDRESS QUESTION---")
|
308 |
+
return "not useful"
|
309 |
+
else:
|
310 |
+
for document in fitting_documents:
|
311 |
+
print(document.page_content)
|
312 |
+
print("---DECISION: THOSE DOCUMENTS ARE NOT GROUNDING THIS GENERATION---")
|
313 |
+
print(f"{generation = }")
|
314 |
+
return "not supported"
|
315 |
+
|
316 |
+
def compile_graph():
|
317 |
+
workflow = StateGraph(GraphState)
|
318 |
+
# Define the nodes
|
319 |
+
workflow.add_node("dora_rewrite", dora_rewrite) # retrieve
|
320 |
+
workflow.add_node("retrieve", retrieve) # retrieve
|
321 |
+
workflow.add_node("grade_documents", grade_documents) # grade documents
|
322 |
+
workflow.add_node("generate", generate) # generate
|
323 |
+
workflow.add_node("transform_query", transform_query) # transform_query
|
324 |
+
# Define the edges
|
325 |
+
workflow.add_edge(START, "dora_rewrite")
|
326 |
+
workflow.add_conditional_edges(
|
327 |
+
"dora_rewrite",
|
328 |
+
suitable_question,
|
329 |
+
{
|
330 |
+
"retrieve": "retrieve",
|
331 |
+
"end": END,
|
332 |
+
},
|
333 |
+
)
|
334 |
+
workflow.add_edge("retrieve", "grade_documents")
|
335 |
+
workflow.add_conditional_edges(
|
336 |
+
"grade_documents",
|
337 |
+
decide_to_generate,
|
338 |
+
{
|
339 |
+
"transform_query": "transform_query",
|
340 |
+
"generate": "generate",
|
341 |
+
},
|
342 |
+
)
|
343 |
+
workflow.add_edge("transform_query", "retrieve")
|
344 |
+
workflow.add_conditional_edges(
|
345 |
+
"generate",
|
346 |
+
grade_generation_v_documents_and_question,
|
347 |
+
{
|
348 |
+
"not supported": "generate",
|
349 |
+
"useful": END,
|
350 |
+
"not useful": "transform_query",
|
351 |
+
},
|
352 |
+
)
|
353 |
+
# Compile
|
354 |
+
app = workflow.compile()
|
355 |
+
return app
|
356 |
+
|
357 |
+
# Function to interact with Gradio
|
358 |
+
def generate_response(question: str, dora: bool, rts: bool, news: bool):
|
359 |
+
selected_sources = [dora, rts, news] if any([dora, rts, news]) else [True, False, False]
|
360 |
+
state = app.invoke({"question": question, "selected_sources": selected_sources})
|
361 |
+
return (
|
362 |
+
state["generation"],
|
363 |
+
('\n\n'.join([f"***{doc.metadata["source"]} section {doc.metadata['section']}***: {doc.page_content}" for doc in state["dora_docs"]])) if "dora_docs" in state and state["dora_docs"] else 'No documents available.',
|
364 |
+
('\n\n'.join([f"***{doc.metadata["source"]}, section {doc.metadata['section']}***: {doc.page_content}" for doc in state["dora_rts_docs"]])) if "dora_rts_docs" in state and state["dora_rts_docs"] else 'No documents available.',
|
365 |
+
('\n\n'.join([f"***{doc.metadata["source"]}***: {doc.page_content}" for doc in state["dora_news_docs"]])) if "dora_news_docs" in state and state["dora_news_docs"] else 'No documents available.',
|
366 |
+
)
|
367 |
+
|
368 |
+
def show_loading(prompt: str):
|
369 |
+
return [prompt, "loading", "loading", "loading", "loading"]
|
370 |
+
|
371 |
+
def on_click():
|
372 |
+
return "I would love to hear your opinion: \[email protected]"
|
373 |
+
|
374 |
+
def clear_results():
|
375 |
+
return "", "", "", "", ""
|
376 |
+
|
377 |
+
def random_prompt():
|
378 |
+
return random.choice([
|
379 |
+
"Was ist der Unterschied zwischen TIBER-EU und DORA TLPT?",
|
380 |
+
"Ich möchte ein SIEM einführen. Bitte gib mir eine Checkliste, was ich beachten muss.",
|
381 |
+
"Was ist der Geltungsbereich der DORA? Bin ich als Finanzdienstleister im Leasinggeschäft betroffen?",
|
382 |
+
"Ich hatte einen Ransomwarevorfall mit erheblichen Auswirkungen auf den Geschäftsbetrieb. Muss ich etwas melden?",
|
383 |
+
"Was ist dieses DORA überhaupt?"
|
384 |
+
])
|
385 |
+
|
386 |
+
def load_css():
|
387 |
+
with open('style.css', 'r') as file:
|
388 |
+
return file.read()
|
389 |
+
|
390 |
+
def run_gradio():
|
391 |
+
with gr.Blocks(title='Artificial Compliance', theme=gr.themes.Monochrome(), css=load_css(), fill_width=True, fill_height=True,) as gradio_ui:
|
392 |
+
# Adding a sliding navbar
|
393 |
+
with gr.Column(scale=1, elem_id='navbar'):
|
394 |
+
gr.Image(
|
395 |
+
'..\\deployment\\logo.png',
|
396 |
+
interactive=False,
|
397 |
+
show_label=False,
|
398 |
+
scale=1,
|
399 |
+
width="50%",
|
400 |
+
height="50%"
|
401 |
+
)
|
402 |
+
with gr.Column():
|
403 |
+
dora_chatbot_button = gr.Checkbox(label="Dora", value=True, elem_classes=["navbar-button"])
|
404 |
+
document_workbench_button = gr.Checkbox(label="Published RTS documents", value=True, elem_classes=["navbar-button"])
|
405 |
+
newsfeed_button = gr.Checkbox(label="Bafin documents", value=True, elem_classes=["navbar-button"])
|
406 |
+
question_prompt = gr.Textbox(
|
407 |
+
value=random_prompt(),
|
408 |
+
label='What you always wanted to know about Dora:',
|
409 |
+
elem_classes=['textbox'],
|
410 |
+
lines=6
|
411 |
+
)
|
412 |
+
with gr.Row():
|
413 |
+
clear_results_button = gr.Button('Clear Results', variant='secondary', size="m")
|
414 |
+
submit_button = gr.Button('Submit', variant='primary', size="m")
|
415 |
+
|
416 |
+
# Adding a header
|
417 |
+
gr.Markdown("# The Doracle", elem_id="header")
|
418 |
+
gr.Markdown("----------------------------------------------------------------------------")
|
419 |
+
display_prompt = gr.Markdown(
|
420 |
+
value="",
|
421 |
+
label="question_prompt",
|
422 |
+
elem_id="header"
|
423 |
+
)
|
424 |
+
gr.Markdown("----------------------------------------------------------------------------")
|
425 |
+
|
426 |
+
with gr.Column(scale=1):
|
427 |
+
with gr.Row(elem_id='text_block'):
|
428 |
+
llm_generation = gr.Markdown(label="LLM Generation", elem_id="llm_generation")
|
429 |
+
|
430 |
+
gr.Markdown("----------------------------------------------------------------------------")
|
431 |
+
|
432 |
+
with gr.Row(elem_id='text_block'):
|
433 |
+
dora_documents = gr.Markdown(label="DORA Documents")
|
434 |
+
dora_rts_documents = gr.Markdown(label="DORA RTS Documents")
|
435 |
+
dora_news_documents = gr.Markdown(label="Bafin supporting Documents")
|
436 |
+
|
437 |
+
# Adding a footer with impressum and contact
|
438 |
+
with gr.Row(elem_classes="footer"):
|
439 |
+
gr.Markdown("Contact", elem_id="clickable_markdown")
|
440 |
+
invisible_btn = gr.Button("", elem_id="invisible_button")
|
441 |
+
|
442 |
+
gr.on(
|
443 |
+
triggers=[question_prompt.submit, submit_button.click],
|
444 |
+
inputs=[question_prompt],
|
445 |
+
outputs=[display_prompt, llm_generation, dora_documents, dora_rts_documents, dora_news_documents],
|
446 |
+
fn=show_loading
|
447 |
+
).then(
|
448 |
+
outputs=[llm_generation, dora_documents, dora_rts_documents, dora_news_documents],
|
449 |
+
inputs=[question_prompt, dora_chatbot_button, document_workbench_button, newsfeed_button],
|
450 |
+
fn=generate_response
|
451 |
+
)
|
452 |
+
|
453 |
+
# Use gr.on() with the invisible button's click event
|
454 |
+
gr.on(
|
455 |
+
triggers=[invisible_btn.click],
|
456 |
+
fn=on_click,
|
457 |
+
outputs=[llm_generation]
|
458 |
+
)
|
459 |
+
|
460 |
+
# Clearing out all results when the appropriate button is clicked
|
461 |
+
clear_results_button.click(fn=clear_results, outputs=[display_prompt, llm_generation, dora_documents, dora_rts_documents, dora_news_documents])
|
462 |
+
|
463 |
+
gradio_ui.launch()
|
464 |
+
|
465 |
+
|
466 |
+
if __name__ == "__main__":
|
467 |
+
_set_env("OPENAI_API_KEY")
|
468 |
+
set_llm_cache(SQLiteCache(database_path=".cache.db"))
|
469 |
+
|
470 |
+
dora_retriever, dora_rts_retriever, dora_news_retriever = load_vectorstores(
|
471 |
+
["./dora_vectorstore_data_faiss.vst",
|
472 |
+
"./rts_eur_lex_vectorstore_faiss.vst",
|
473 |
+
"./bafin_news_vectorstore_faiss.vst",]
|
474 |
+
)
|
475 |
+
|
476 |
+
fast_llm = ChatOpenAI(model="gpt-3.5-turbo")
|
477 |
+
smart_llm = ChatOpenAI(model="gpt-4-turbo", temperature=0.2, max_tokens=4096)
|
478 |
+
tool_llm = ChatOpenAI(model="gpt-4o")
|
479 |
+
rewrite_llm = ChatOpenAI(model="gpt-3.5-turbo", temperature=1, cache=False)
|
480 |
+
|
481 |
+
dora_question_rewriter = IMPROVE_PROMPT | tool_llm | StrOutputParser()
|
482 |
+
retrieval_grader = RELEVANCE_PROMPT | fast_llm.with_structured_output(GradeDocuments)
|
483 |
+
answer_chain = ANSWER_PROMPT | smart_llm | StrOutputParser() #former RAG chain
|
484 |
+
hallucination_grader = HALLUCINATION_PROMPT | fast_llm.with_structured_output(GradeHallucinations)
|
485 |
+
answer_grader = RESOLVER_PROMPT | fast_llm.with_structured_output(GradeAnswer)
|
486 |
+
question_rewriter = REWRITER_PROMPT | rewrite_llm | StrOutputParser()
|
487 |
+
|
488 |
+
app = compile_graph()
|
489 |
+
|
490 |
+
# And finally, run the app
|
491 |
+
run_gradio()
|
492 |
+
|
493 |
+
|
494 |
+
|
logo.png
ADDED
prompts.py
ADDED
@@ -0,0 +1,69 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
from langchain_core.prompts import ChatPromptTemplate
|
2 |
+
|
3 |
+
# Rewriting process using ChatPromptTemplate
|
4 |
+
IMPROVE_PROMPT = ChatPromptTemplate.from_messages(
|
5 |
+
[
|
6 |
+
("system", """
|
7 |
+
You are a question rewriter that optimizes an input question for better retrieval from vectorstore data containing information security regulatory texts, especially tha digital operations security acrt (DORA).
|
8 |
+
The regulatory texts in the vectorstore mainly address the following topics: {topics}.
|
9 |
+
Your goal is to understand the underlying semantic intent of the input question and reformulate it to improve clarity and relevance for retrieving content on {topics}.
|
10 |
+
If the question is not related to any of the topics or information security regulatory texts, simply answer: "Thats an interesting question, but I dont think I can answer it based on my Dora knowledge."
|
11 |
+
"""),
|
12 |
+
(
|
13 |
+
"human",
|
14 |
+
"Here is the initial question: \n\n{question} \nPlease formulate an improved version of the question, focusing on clarity and retrieval optimization."
|
15 |
+
),
|
16 |
+
]
|
17 |
+
)
|
18 |
+
|
19 |
+
RELEVANCE_PROMPT = ChatPromptTemplate.from_messages(
|
20 |
+
[
|
21 |
+
("system", """You are a grader assessing relevance of a retrieved document to a user question. \n
|
22 |
+
If the document contains keyword(s) or semantic meaning related to the user question, grade it as relevant. \n
|
23 |
+
It does not need to be a stringent test. The goal is to filter out erroneous retrievals. \n
|
24 |
+
Give a binary score 'yes' or 'no' score to indicate whether the document is relevant to the question."""
|
25 |
+
),
|
26 |
+
("human", "Retrieved document: \n\n {document} \n\n User question: {question}"),
|
27 |
+
]
|
28 |
+
)
|
29 |
+
|
30 |
+
ANSWER_PROMPT = ChatPromptTemplate.from_messages(
|
31 |
+
[
|
32 |
+
(
|
33 |
+
"system",
|
34 |
+
"You are a highly experienced IT auditor, specializing in information security and regulatory compliance. Your task is to assist a colleague who has approached you with a question."
|
35 |
+
" You have access to relevant context, provided here: {context}."
|
36 |
+
" Please respond with a clear, concise, and precise answer, strictly based on the provided context. Ensure your response is accurate and always cite sources from the context."
|
37 |
+
" Do not introduce any new information or alter the context in any way."
|
38 |
+
),
|
39 |
+
("user", "{question}"),
|
40 |
+
]
|
41 |
+
)
|
42 |
+
|
43 |
+
HALLUCINATION_PROMPT = ChatPromptTemplate.from_messages(
|
44 |
+
|
45 |
+
[
|
46 |
+
("system", """You are a grader assessing whether an LLM generation is grounded in / supported by a set of retrieved facts. \n
|
47 |
+
Give a binary score 'yes' or 'no'. 'Yes' means that the answer is grounded in / supported by the set of facts."""),
|
48 |
+
("human", "Set of facts: \n\n {documents} \n\n LLM generation: {generation}"),
|
49 |
+
]
|
50 |
+
)
|
51 |
+
|
52 |
+
RESOLVER_PROMPT = ChatPromptTemplate.from_messages(
|
53 |
+
[
|
54 |
+
("system", """You are a grader assessing whether an answer addresses / resolves a question \n
|
55 |
+
Give a binary score 'yes' or 'no'. Yes' means that the answer resolves the question."""),
|
56 |
+
("human", "User question: \n\n {question} \n\n LLM generation: {generation}"),
|
57 |
+
]
|
58 |
+
)
|
59 |
+
|
60 |
+
REWRITER_PROMPT = ChatPromptTemplate.from_messages(
|
61 |
+
[
|
62 |
+
("system", """You a question re-writer that converts an input question to a better version that is optimized \n
|
63 |
+
for vectorstore retrieval. Look at the input and try to reason about the underlying semantic intent / meaning."""),
|
64 |
+
(
|
65 |
+
"human",
|
66 |
+
"Here is the initial question: \n\n {question} \n Formulate an improved question.",
|
67 |
+
),
|
68 |
+
]
|
69 |
+
)
|
requirements.txt
ADDED
@@ -0,0 +1,7 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
gradio==5.4.0
|
2 |
+
langchain_community==0.3.4
|
3 |
+
langchain_core==0.3.14
|
4 |
+
langchain_openai==0.2.4
|
5 |
+
langgraph==0.2.41
|
6 |
+
pydantic==2.9.2
|
7 |
+
typing_extensions==4.12.2
|
style.css
ADDED
@@ -0,0 +1,65 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
/* style.css */
|
2 |
+
.gradio-container {
|
3 |
+
width: 100% !important;
|
4 |
+
max-width: none !important;
|
5 |
+
background-color: #D3D3D3;
|
6 |
+
}
|
7 |
+
#navbar {
|
8 |
+
position: fixed;
|
9 |
+
left: -190px;
|
10 |
+
top: 0;
|
11 |
+
width: 200px;
|
12 |
+
height: 100%;
|
13 |
+
background-color: #444;
|
14 |
+
transition: left 0.3s ease;
|
15 |
+
padding: 10px;
|
16 |
+
box-shadow: 2px 0 5px rgba(0, 0, 0, 0.5);
|
17 |
+
z-index: 1000;
|
18 |
+
}
|
19 |
+
#navbar:hover {
|
20 |
+
left: 0;
|
21 |
+
}
|
22 |
+
.navbar-button {
|
23 |
+
background-color: #D3D3D3;
|
24 |
+
color: #fff;
|
25 |
+
margin-bottom: 10px;
|
26 |
+
width: 100%;
|
27 |
+
}
|
28 |
+
.footer {
|
29 |
+
text-align: center;
|
30 |
+
padding: 10px;
|
31 |
+
background-color: #ffffff;
|
32 |
+
margin-top: 20px;
|
33 |
+
}
|
34 |
+
#header {
|
35 |
+
text-align: center;
|
36 |
+
font-size: 1.8em;
|
37 |
+
font-weight: bold;
|
38 |
+
margin-top: 20px;
|
39 |
+
}
|
40 |
+
#text_block {
|
41 |
+
margin-left: 150px;
|
42 |
+
margin-right: 150px;
|
43 |
+
max-width: calc(100% - 300px);
|
44 |
+
box-sizing: border-box;
|
45 |
+
overflow-x: hidden;
|
46 |
+
word-wrap: break-word;
|
47 |
+
}
|
48 |
+
#llm_output {
|
49 |
+
width: 100%;
|
50 |
+
box-sizing: border-box;
|
51 |
+
padding: 10px;
|
52 |
+
}
|
53 |
+
#clickable_markdown {
|
54 |
+
position: relative;
|
55 |
+
cursor: pointer;
|
56 |
+
}
|
57 |
+
#invisible_button {
|
58 |
+
position: absolute;
|
59 |
+
top: 0;
|
60 |
+
left: 0;
|
61 |
+
width: 100%;
|
62 |
+
height: 100%;
|
63 |
+
opacity: 0;
|
64 |
+
cursor: pointer;
|
65 |
+
}
|