Spaces:
Running
Running
Create retriever/llm_manager.py
Browse files- retriever/llm_manager.py +309 -0
retriever/llm_manager.py
ADDED
@@ -0,0 +1,309 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import logging
|
2 |
+
import os
|
3 |
+
from typing import List, Dict, Any, Tuple
|
4 |
+
from langchain_groq import ChatGroq
|
5 |
+
from langchain.chains import RetrievalQA
|
6 |
+
from langchain_core.documents import Document
|
7 |
+
from langchain_core.retrievers import BaseRetriever
|
8 |
+
from langchain.chains.summarize import load_summarize_chain
|
9 |
+
from langchain.prompts import PromptTemplate
|
10 |
+
|
11 |
+
class LLMManager:
|
12 |
+
DEFAULT_MODEL = "gemma2-9b-it" # Set the default model name
|
13 |
+
|
14 |
+
def __init__(self):
|
15 |
+
self.generation_llm = None
|
16 |
+
logging.info("LLMManager initialized")
|
17 |
+
|
18 |
+
# Initialize the default model during construction
|
19 |
+
try:
|
20 |
+
self.initialize_generation_llm(self.DEFAULT_MODEL)
|
21 |
+
logging.info(f"Initialized default LLM model: {self.DEFAULT_MODEL}")
|
22 |
+
except ValueError as e:
|
23 |
+
logging.error(f"Failed to initialize default LLM model: {str(e)}")
|
24 |
+
|
25 |
+
def initialize_generation_llm(self, model_name: str) -> None:
|
26 |
+
"""
|
27 |
+
Initialize the generation LLM using the Groq API.
|
28 |
+
|
29 |
+
Args:
|
30 |
+
model_name (str): The name of the model to use for generation.
|
31 |
+
|
32 |
+
Raises:
|
33 |
+
ValueError: If GROQ_API_KEY is not set.
|
34 |
+
"""
|
35 |
+
api_key = os.getenv("GROQ_API_KEY")
|
36 |
+
if not api_key:
|
37 |
+
raise ValueError("GROQ_API_KEY is not set. Please add it in your environment variables.")
|
38 |
+
|
39 |
+
os.environ["GROQ_API_KEY"] = api_key
|
40 |
+
self.generation_llm = ChatGroq(model=model_name, temperature=0.7)
|
41 |
+
self.generation_llm.name = model_name
|
42 |
+
logging.info(f"Generation LLM {model_name} initialized")
|
43 |
+
|
44 |
+
def reinitialize_llm(self, model_name: str) -> str:
|
45 |
+
"""
|
46 |
+
Reinitialize the LLM with a new model name.
|
47 |
+
|
48 |
+
Args:
|
49 |
+
model_name (str): The name of the new model to initialize.
|
50 |
+
|
51 |
+
Returns:
|
52 |
+
str: Status message indicating success or failure.
|
53 |
+
"""
|
54 |
+
try:
|
55 |
+
self.initialize_generation_llm(model_name)
|
56 |
+
return f"LLM model changed to {model_name}"
|
57 |
+
except ValueError as e:
|
58 |
+
logging.error(f"Failed to reinitialize LLM with model {model_name}: {str(e)}")
|
59 |
+
return f"Error: Failed to change LLM model: {str(e)}"
|
60 |
+
|
61 |
+
def generate_response(self, question: str, relevant_docs: List[Dict[str, Any]]) -> Tuple[str, List[Document]]:
|
62 |
+
"""
|
63 |
+
Generate a response using the generation LLM based on the question and relevant documents.
|
64 |
+
|
65 |
+
Args:
|
66 |
+
question (str): The user's query.
|
67 |
+
relevant_docs (List[Dict[str, Any]]): List of relevant document chunks with text, metadata, and scores.
|
68 |
+
|
69 |
+
Returns:
|
70 |
+
Tuple[str, List[Document]]: The LLM's response and the source documents used.
|
71 |
+
|
72 |
+
Raises:
|
73 |
+
ValueError: If the generation LLM is not initialized.
|
74 |
+
Exception: If there's an error during the QA chain invocation.
|
75 |
+
"""
|
76 |
+
if not self.generation_llm:
|
77 |
+
raise ValueError("Generation LLM is not initialized. Call initialize_generation_llm first.")
|
78 |
+
|
79 |
+
# Convert the relevant documents into LangChain Document objects
|
80 |
+
documents = [
|
81 |
+
Document(page_content=doc['text'], metadata=doc['metadata'])
|
82 |
+
for doc in relevant_docs
|
83 |
+
]
|
84 |
+
|
85 |
+
# Create a proper retriever by subclassing BaseRetriever
|
86 |
+
class SimpleRetriever(BaseRetriever):
|
87 |
+
def __init__(self, docs: List[Document], **kwargs):
|
88 |
+
super().__init__(**kwargs) # Pass kwargs to BaseRetriever
|
89 |
+
self._docs = docs # Use a private attribute to store docs
|
90 |
+
logging.debug(f"SimpleRetriever initialized with {len(docs)} documents")
|
91 |
+
|
92 |
+
def _get_relevant_documents(self, query: str) -> List[Document]:
|
93 |
+
logging.debug(f"SimpleRetriever._get_relevant_documents called with query: {query}")
|
94 |
+
return self._docs
|
95 |
+
|
96 |
+
async def _aget_relevant_documents(self, query: str) -> List[Document]:
|
97 |
+
logging.debug(f"SimpleRetriever._aget_relevant_documents called with query: {query}")
|
98 |
+
return self._docs
|
99 |
+
|
100 |
+
# Instantiate the retriever
|
101 |
+
retriever = SimpleRetriever(docs=documents)
|
102 |
+
|
103 |
+
# Create a retrieval-based question-answering chain
|
104 |
+
qa_chain = RetrievalQA.from_chain_type(
|
105 |
+
llm=self.generation_llm,
|
106 |
+
retriever=retriever,
|
107 |
+
return_source_documents=True
|
108 |
+
)
|
109 |
+
|
110 |
+
try:
|
111 |
+
result = qa_chain.invoke({"query": question})
|
112 |
+
response = result['result']
|
113 |
+
source_docs = result['source_documents']
|
114 |
+
#logging.info(f"Generated response for question: {question} : {response}")
|
115 |
+
return response, source_docs
|
116 |
+
except Exception as e:
|
117 |
+
logging.error(f"Error during QA chain invocation: {str(e)}")
|
118 |
+
raise e
|
119 |
+
|
120 |
+
def generate_summary_v0(self, chunks: any):
|
121 |
+
logging.info("Generating summary ...")
|
122 |
+
|
123 |
+
# Limit the number of chunks (for example, top 30 chunks)
|
124 |
+
limited_chunks = chunks[:30]
|
125 |
+
|
126 |
+
# Combine text from the selected chunks
|
127 |
+
full_text = "\n".join(chunk['text'] for chunk in limited_chunks)
|
128 |
+
text_length = len(full_text)
|
129 |
+
logging.info(f"Total text length (characters): {text_length}")
|
130 |
+
|
131 |
+
# Define a maximum character limit to fit in a 1024-token context.
|
132 |
+
# For many models, roughly 3200 characters is a safe limit.
|
133 |
+
MAX_CHAR_LIMIT = 3200
|
134 |
+
if text_length > MAX_CHAR_LIMIT:
|
135 |
+
logging.warning(f"Input text too long ({text_length} chars), truncating to {MAX_CHAR_LIMIT} chars.")
|
136 |
+
full_text = full_text[:MAX_CHAR_LIMIT]
|
137 |
+
|
138 |
+
# Define a custom prompt to instruct concise summarization in bullet points.
|
139 |
+
custom_prompt_template = """
|
140 |
+
You are an expert summarizer. Summarize the following text into a concise summary using bullet points.
|
141 |
+
Ensure that the final summary is no longer than 20-30 bullet points and fits within 15-20 lines.
|
142 |
+
Focus only on the most critical points.
|
143 |
+
|
144 |
+
Text to summarize:
|
145 |
+
{text}
|
146 |
+
|
147 |
+
Summary:
|
148 |
+
"""
|
149 |
+
prompt = PromptTemplate(input_variables=["text"], template=custom_prompt_template)
|
150 |
+
|
151 |
+
# Use the 'stuff' chain type to send a single LLM request with our custom prompt.
|
152 |
+
chain = load_summarize_chain(self.generation_llm, chain_type="stuff", prompt=prompt)
|
153 |
+
|
154 |
+
# Wrap the full text in a single Document object (chain expects a list of Documents)
|
155 |
+
docs = [Document(page_content=full_text)]
|
156 |
+
|
157 |
+
# Generate the summary
|
158 |
+
summary = chain.invoke(docs)
|
159 |
+
return summary['output_text']
|
160 |
+
|
161 |
+
def generate_questions(self, chunks: any):
|
162 |
+
logging.info("Generating sample questions ...")
|
163 |
+
|
164 |
+
# Use the top 30 chunks or fewer
|
165 |
+
limited_chunks = chunks[:30]
|
166 |
+
|
167 |
+
# Combine text from chunks
|
168 |
+
full_text = "\n".join(chunk['text'] for chunk in limited_chunks)
|
169 |
+
text_length = len(full_text)
|
170 |
+
logging.info(f"Total text length for questions: {text_length}")
|
171 |
+
|
172 |
+
MAX_CHAR_LIMIT = 3200
|
173 |
+
if text_length > MAX_CHAR_LIMIT:
|
174 |
+
logging.warning(f"Input text too long ({text_length} chars), truncating to {MAX_CHAR_LIMIT} chars.")
|
175 |
+
full_text = full_text[:MAX_CHAR_LIMIT]
|
176 |
+
|
177 |
+
# Prompt template for generating questions
|
178 |
+
question_prompt_template = """
|
179 |
+
You are an AI expert at creating questions from documents.
|
180 |
+
|
181 |
+
Based on the text below, generate not less than 20 insightful and highly relevant sample questions that a user might ask to better understand the content.
|
182 |
+
|
183 |
+
**Instructions:**
|
184 |
+
- Questions must be specific to the document's content and context.
|
185 |
+
- Avoid generic questions like 'What is this document about?'
|
186 |
+
- Do not include numbers, prefixes (e.g., '1.', '2.'), or explanations (e.g., '(Clarifies...)').
|
187 |
+
- Each question should be a single, clear sentence ending with a question mark.
|
188 |
+
- Focus on key concepts, processes, components, or use cases mentioned in the text.
|
189 |
+
|
190 |
+
Text:
|
191 |
+
{text}
|
192 |
+
|
193 |
+
Output format:
|
194 |
+
What is the purpose of the Communication Server in Collateral Management?
|
195 |
+
How does the system handle data encryption for secure communication?
|
196 |
+
...
|
197 |
+
"""
|
198 |
+
prompt = PromptTemplate(input_variables=["text"], template=question_prompt_template)
|
199 |
+
|
200 |
+
chain = load_summarize_chain(self.generation_llm, chain_type="stuff", prompt=prompt)
|
201 |
+
docs = [Document(page_content=full_text)]
|
202 |
+
|
203 |
+
try:
|
204 |
+
result = chain.invoke(docs)
|
205 |
+
question_output = result.get("output_text", "").strip()
|
206 |
+
|
207 |
+
# Clean and parse the output into a list of questions
|
208 |
+
questions = []
|
209 |
+
for line in question_output.split("\n"):
|
210 |
+
# Remove any leading/trailing whitespace, numbers, or bullet points
|
211 |
+
cleaned_line = line.strip().strip("-*1234567890. ").rstrip(".")
|
212 |
+
# Remove any explanation in parentheses
|
213 |
+
cleaned_line = cleaned_line.split("(")[0].strip()
|
214 |
+
# Ensure the line is a valid question (ends with '?' and is not empty)
|
215 |
+
if cleaned_line and cleaned_line.endswith("?"):
|
216 |
+
questions.append(cleaned_line)
|
217 |
+
|
218 |
+
# Limit to 10 questions
|
219 |
+
questions = questions[:10]
|
220 |
+
logging.info(f"Generated questions: {questions}")
|
221 |
+
return questions
|
222 |
+
except Exception as e:
|
223 |
+
logging.error(f"Error generating questions: {e}")
|
224 |
+
return []
|
225 |
+
|
226 |
+
def generate_summary(self, chunks: Any, toc_text: Any, summary_type: str = "medium") -> str:
|
227 |
+
"""
|
228 |
+
Generate a summary of the document using LangChain's summarization chains.
|
229 |
+
|
230 |
+
Args:
|
231 |
+
vector_store_manager: Instance of VectorStoreManager with a FAISS vector store.
|
232 |
+
summary_type (str): Type of summary ("small", "medium", "detailed").
|
233 |
+
k (int): Number of chunks to retrieve from the vector store.
|
234 |
+
include_toc (bool): Whether to include the table of contents (if available).
|
235 |
+
|
236 |
+
Returns:
|
237 |
+
str: Generated summary.
|
238 |
+
|
239 |
+
Raises:
|
240 |
+
ValueError: If summary_type is invalid or vector store is not initialized.
|
241 |
+
"""
|
242 |
+
|
243 |
+
# Define chunk retrieval parameters based on summary type
|
244 |
+
if summary_type == "small":
|
245 |
+
k = min(k, 3) # Fewer chunks for small summary
|
246 |
+
chain_type = "stuff" # Use stuff for small summaries
|
247 |
+
word_count = "50-100"
|
248 |
+
elif summary_type == "medium":
|
249 |
+
k = min(k, 10)
|
250 |
+
chain_type = "map_reduce" # Use map-reduce for medium summaries
|
251 |
+
word_count = "200-400"
|
252 |
+
else: # detailed
|
253 |
+
k = min(k, 20)
|
254 |
+
chain_type = "map_reduce" # Use map-reduce for detailed summaries
|
255 |
+
word_count = "500-1000"
|
256 |
+
|
257 |
+
# Define prompts
|
258 |
+
if chain_type == "stuff":
|
259 |
+
prompt = PromptTemplate(
|
260 |
+
input_variables=["text"],
|
261 |
+
template=(
|
262 |
+
"Generate a {summary_type} summary ({word_count} words) of the following document excerpts. "
|
263 |
+
"Focus on key points and ensure clarity. Stick strictly to the provided text:\n\n"
|
264 |
+
"{toc_prompt}{text}"
|
265 |
+
).format(
|
266 |
+
summary_type=summary_type,
|
267 |
+
word_count=word_count,
|
268 |
+
toc_prompt="Table of Contents:\n{toc_text}\n\n" if toc_text else ""
|
269 |
+
)
|
270 |
+
)
|
271 |
+
chain = load_summarize_chain(
|
272 |
+
llm=self.generation_llm,
|
273 |
+
chain_type="stuff",
|
274 |
+
prompt=prompt
|
275 |
+
)
|
276 |
+
else: # map_reduce
|
277 |
+
map_prompt = PromptTemplate(
|
278 |
+
input_variables=["text"],
|
279 |
+
template=(
|
280 |
+
"Summarize the following document excerpt in 1-2 sentences, focusing on key points. "
|
281 |
+
"Consider the document's structure from this table of contents:\n\n"
|
282 |
+
"Table of Contents:\n{toc_text}\n\nExcerpt:\n{text}"
|
283 |
+
).format(toc_text=toc_text if toc_text else "Not provided")
|
284 |
+
)
|
285 |
+
combine_prompt = PromptTemplate(
|
286 |
+
input_variables=["text"],
|
287 |
+
template=(
|
288 |
+
"Combine the following summaries into a cohesive {summary_type} summary "
|
289 |
+
"({word_count} words) of the document. Ensure clarity, avoid redundancy, and "
|
290 |
+
"organize by key themes or sections if applicable:\n\n{text}"
|
291 |
+
).format(summary_type=summary_type, word_count=word_count)
|
292 |
+
)
|
293 |
+
chain = load_summarize_chain(
|
294 |
+
llm=self.generation_llm,
|
295 |
+
chain_type="map_reduce",
|
296 |
+
map_prompt=map_prompt,
|
297 |
+
combine_prompt=combine_prompt,
|
298 |
+
return_intermediate_steps=False
|
299 |
+
)
|
300 |
+
|
301 |
+
# Run the chain
|
302 |
+
try:
|
303 |
+
logging.info(f"Generating {summary_type} summary with {len(chunks)} chunks")
|
304 |
+
summary = chain.run(chunks)
|
305 |
+
logging.info(f"{summary_type.capitalize()} summary generated successfully")
|
306 |
+
return summary
|
307 |
+
except Exception as e:
|
308 |
+
logging.error(f"Error generating summary: {str(e)}")
|
309 |
+
return f"Error generating summary: {str(e)}"
|