Add context compression mechanism

This commit is contained in:
2025-02-09 12:05:41 +08:00
parent 32e61b8d24
commit 6a0cab65d1
4 changed files with 76 additions and 41 deletions

View File

@@ -2,17 +2,23 @@ import os
import re
import time
import uuid
import tiktoken
import xml.etree.ElementTree as ET
from langchain_openai import ChatOpenAI
from langchain_openai import OpenAIEmbeddings
from langchain_openai import ChatOpenAI, OpenAIEmbeddings
from langchain_core.messages import SystemMessage
from langchain_core.runnables.history import RunnableWithMessageHistory
from langchain_core.prompts import ChatPromptTemplate, MessagesPlaceholder
from langchain_community.vectorstores import FAISS
from langchain_community.chat_message_histories import ChatMessageHistory
from langchain_community.document_transformers import EmbeddingsRedundantFilter
from langchain.retrievers import ContextualCompressionRetriever
from langchain.retrievers.document_compressors import EmbeddingsFilter, DocumentCompressorPipeline
from langchain_text_splitters import CharacterTextSplitter
from logger import Logger
from audit import callback
from audit.prompt import SYSTEM_PROMPT
from audit.language import LANGUAGE
reasoning_model = 'gemini-2.0-flash-thinking-exp'
embedding_model = 'text-embedding-3-large'
@@ -24,12 +30,36 @@ class Audit:
def __init__(self):
self.raw_chain = None
self.source_files_list = []
self.max_token = 4096
self.chat_history = ChatMessageHistory()
self.session_id = uuid.uuid4().hex
self.response_callback = callback.CustomCallbackHandler()
self.embedding = OpenAIEmbeddings(model=embedding_model)
self.llm = ChatOpenAI(model=reasoning_model, streaming=True, callbacks=[self.response_callback])
self.llm = ChatOpenAI(
model=reasoning_model,
streaming=True,
callbacks=[self.response_callback]
)
self.log = Logger('audit')
self.splitter = CharacterTextSplitter(
chunk_size=300,
chunk_overlap=0,
separator=". ",
)
self.messages_db = FAISS.from_texts(['nothing'], self.embedding)
self.retriever = self.messages_db.as_retriever()
self.redundant_filter = EmbeddingsRedundantFilter(embeddings=self.embedding)
self.relevant_filter = EmbeddingsFilter(
embeddings=self.embedding,
similarity_threshold=0.76,
)
self.pipeline_compressor = DocumentCompressorPipeline(
transformers=[self.splitter, self.redundant_filter, self.relevant_filter]
)
self.compression_retriever = ContextualCompressionRetriever(
base_compressor=self.pipeline_compressor,
base_retriever=self.retriever,
)
self.prompt = ChatPromptTemplate.from_messages([
SystemMessage(content=SYSTEM_PROMPT),
MessagesPlaceholder(variable_name='messages'),
@@ -41,19 +71,22 @@ class Audit:
input_content = ''
while True:
time.sleep(3)
result = self.send_message(input_content)
xml_match = re.search(xml_pattern, result, re.DOTALL)
if xml_match:
xml_content = xml_match.group(0)
root = ET.fromstring(xml_content)
if xml_match := re.search(xml_pattern, result, re.DOTALL):
try:
xml_content = xml_match.group(0)
root = ET.fromstring(xml_content)
action = root.find('action').text
content = root.find('content').text
action = root.find('action').text
content = root.find('content').text
except Exception as e:
self.log.error(f'Illegal output, try to correct')
input_content = 'ILLEGAL OUTPUT'
continue
if action == 'QUERY STRUCTURE':
self.log.info('Request to query project structure')
self.log.info('Request project structure')
input_content = '\n'.join(x for x in self.source_files_list)
continue
elif action == 'QUERY SOURCE':
@@ -61,7 +94,8 @@ class Audit:
input_content = open(content, 'r', encoding='utf-8').read()
continue
elif action == 'OUTPUT RESULT':
self.log.warning(f'Audit result: \n{content}\n')
self.log.warning(f'Audit result: \n\n{content}')
self.store_messages_in_faiss(content)
callback_function(content) # Callback function, used to obtain results externally
input_content = ''
continue
@@ -71,13 +105,14 @@ class Audit:
def send_message(self, input_content):
self.response_callback.temp_content = ''
compressed_context = self.compression_retriever.invoke(input_content)
if input_content == '':
input_content = 'nothing'
input_dict = {
'input': input_content,
'context': '',
'context': compressed_context,
}
config_dict = {
'configurable': {'session_id': self.session_id}
@@ -91,41 +126,31 @@ class Audit:
history_messages_key='messages',
)
self.log.debug(f'Chat messages: {input_dict}')
for _ in chain_with_message_history.stream(input_dict, config_dict):
pass
return self.response_callback.temp_content
def load_source_files(self, path, language):
def store_messages_in_faiss(self, message):
text_embedding = self.embedding.embed_query(message)
doc_id = str(uuid.uuid4())
self.messages_db.add_embeddings([(doc_id, text_embedding)], metadatas=[{"id": doc_id}])
self.log.info(f"Audit result stored in messages_db with ID: {doc_id}")
def load_source_files(self, path, lang):
self.log.info('Loading source files')
if language == 'php':
suffixes = ['.php', '.php3', 'php4', 'php5']
elif language == 'python':
suffixes = ['.py']
elif language == 'java':
suffixes = ['.java']
elif language == 'c':
suffixes = ['.c']
elif language == 'c++':
suffixes = ['.cpp', 'cc']
elif language == 'javascript':
suffixes = ['.js']
elif language == 'go':
suffixes = ['.go']
if lang in LANGUAGE:
suffixes = LANGUAGE[lang]
else:
self.log.critical('Language not supported!')
return
for root, dirs, files in os.walk(path):
for file_name in files:
hit = False
for suffix in suffixes:
if file_name.endswith(suffix):
hit = True
break
if hit:
self.source_files_list.append(os.path.join(root, file_name))
for root, _, files in os.walk(path):
self.source_files_list.extend(
os.path.join(root, file) for file in files if any(file.endswith(suffix) for suffix in suffixes)
)
self.log.info(f'Finished loading source files. total files: {len(self.source_files_list)}')

9
audit/language.py Normal file
View File

@@ -0,0 +1,9 @@
LANGUAGE = {
'c': ['.c'],
'c++': ['.cc', '.cpp'],
'go': ['.go'],
'php': ['php', 'php3', 'php4', 'php5', 'phtml'],
'java': ['.java'],
'python': ['.py'],
'javascript': ['.js'],
}

View File

@@ -33,4 +33,5 @@ Important things:
3. When you need to query the content of a code file, please note that you can only query one file at a time. Please follow The above format outputs the absolute path of the file to be queried;
4. After you output the audit results, the user will reply with an empty string. Please make sure that all code files have been audited before ending the audit task;
5. In any case, you must strictly follow the several action formats given above for output. Any content outside the output format is prohibited. Do not try to ask or suggest;
6. When the user prompts "ILLEGAL OUTPUT", it means that your output violates the user's specifications. Please confirm again that all your output must comply with the user's specifications.
"""