StanGirard/quivr
1
1
Fork 0
mirror of https://github.com/StanGirard/quivr.git synced 2024-12-24 20:03:41 +03:00
Code Issues Packages Projects Releases Wiki Activity

feat(neurons): added class

Browse Source
This commit is contained in:
Stan Girard 2023-06-19 21:15:35 +02:00
parent d42f14f431
commit dc6f610b26
3 changed files with 45 additions and 49 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

5
backend/parsers/common.py
View File

@ -10,7 +10,7 @@ from langchain.schema import Document
from langchain.text_splitter import RecursiveCharacterTextSplitter
from utils.common import CommonsDep
from utils.file import compute_sha1_from_content, compute_sha1_from_file
from utils.vectors import create_summary, create_vector
from utils.vectors import Neurons, create_summary
async def process_file(commons: CommonsDep, file: UploadFile, loader_class, file_suffix, enable_summarization, user, user_openai_api_key):
@ -52,7 +52,8 @@ async def process_file(commons: CommonsDep, file: UploadFile, loader_class, file
}
doc_with_metadata = Document(
page_content=doc.page_content, metadata=metadata)
create_vector(commons, user.email, doc_with_metadata, user_openai_api_key)
neurons = Neurons(commons=commons)
neurons.create_vector(user.email, doc_with_metadata, user_openai_api_key)
# add_usage(stats_db, "embedding", "audio", metadata={"file_name": file_meta_name,"file_type": ".txt", "chunk_size": chunk_size, "chunk_overlap": chunk_overlap})
# Remove the enable_summarization and ids

5
backend/parsers/github.py
View File

@ -7,7 +7,7 @@ from langchain.text_splitter import RecursiveCharacterTextSplitter
from parsers.common import file_already_exists_from_content
from utils.common import CommonsDep
from utils.file import compute_sha1_from_content
from utils.vectors import create_vector
from utils.vectors import Neurons
async def process_github(commons: CommonsDep, repo, enable_summarization, user, supabase, user_openai_api_key):
@ -44,7 +44,8 @@ async def process_github(commons: CommonsDep, repo, enable_summarization, user,
page_content=doc.page_content, metadata=metadata)
exist = await file_already_exists_from_content(supabase, doc.page_content.encode("utf-8"), user)
if not exist:
create_vector(commons, user.email, doc_with_metadata, user_openai_api_key)
neurons = Neurons(commons=commons)
neurons.create_vector(user.email, doc_with_metadata, user_openai_api_key)
print("Created vector for ", doc.metadata["file_name"])
return {"message": f"✅ Github with {len(documents)} files has been uploaded.", "type": "success"}

84
backend/utils/vectors.py
View File

@ -1,14 +1,43 @@
from langchain.embeddings.openai import OpenAIEmbeddings
from langchain.schema import Document
from llm.qa import BrainPicking
from llm.qa import BrainPicking, BrainSettings
from llm.summarization import llm_evaluate_summaries, llm_summerize
from logger import get_logger
from models.chats import ChatMessage
from pydantic import BaseModel
from utils.common import CommonsDep
logger = get_logger(__name__)
# TO DO: Create classes or other to avoid having to specify commons in each one of these functions
class Neurons(BaseModel):
commons: CommonsDep
settings = BrainSettings()
def create_vector(self, user_id, doc, user_openai_api_key=None):
logger.info(f"Creating vector for document")
logger.info(f"Document: {doc}")
if user_openai_api_key:
self.commons['documents_vector_store']._embedding = OpenAIEmbeddings(openai_api_key=user_openai_api_key)
try:
sids = self.commons['documents_vector_store'].add_documents([doc])
if sids and len(sids) > 0:
self.commons['supabase'].table("vectors").update({"user_id": user_id}).match({"id": sids[0]}).execute()
except Exception as e:
logger.error(f"Error creating vector for document {e}")
def create_embedding(self, content):
return self.commons['embeddings'].embed_query(content)
def similarity_search(self, query, table='match_summaries', top_k=5, threshold=0.5):
query_embedding = self.create_embedding(query)
summaries = self.commons['supabase'].rpc(
table, {'query_embedding': query_embedding,
'match_count': top_k, 'match_threshold': threshold}
).execute()
return summaries.data
def create_summary(commons: CommonsDep, document_id, content, metadata):
logger.info(f"Summarizing document {content[:100]}")
summary = llm_summerize(content)
@ -22,77 +51,42 @@ def create_summary(commons: CommonsDep, document_id, content, metadata):
commons['supabase'].table("summaries").update(
{"document_id": document_id}).match({"id": sids[0]}).execute()
def create_vector(commons: CommonsDep, user_id,doc, user_openai_api_key=None):
logger.info(f"Creating vector for document")
logger.info(f"Document: {doc}")
if user_openai_api_key:
commons['documents_vector_store']._embedding = OpenAIEmbeddings(openai_api_key=user_openai_api_key)
try:
sids = commons['documents_vector_store'].add_documents(
[doc])
if sids and len(sids) > 0:
commons['supabase'].table("vectors").update(
{"user_id": user_id}).match({"id": sids[0]}).execute()
# TODO: create entry in brains_vectors table with brain_id and vector_id
except Exception as e:
logger.error(f"Error creating vector for document {e}")
def create_embedding(commons: CommonsDep, content):
return commons['embeddings'].embed_query(content)
def similarity_search(commons: CommonsDep, query, table='match_summaries', top_k=5, threshold=0.5):
query_embedding = create_embedding(commons, query)
summaries = commons['supabase'].rpc(
table, {'query_embedding': query_embedding,
'match_count': top_k, 'match_threshold': threshold}
).execute()
return summaries.data
def get_answer(commons: CommonsDep, chat_message: ChatMessage, email: str, user_openai_api_key:str):
def get_answer(commons: CommonsDep, chat_message: ChatMessage, email: str, user_openai_api_key: str):
Brain = BrainPicking().init(chat_message.model, email)
qa = Brain.get_qa(chat_message, user_openai_api_key)
neurons = Neurons(commons=commons)
if chat_message.use_summarization:
# 1. get summaries from the vector store based on question
summaries = similarity_search(commons,
chat_message.question, table='match_summaries')
# 2. evaluate summaries against the question
summaries = neurons.similarity_search(chat_message.question, table='match_summaries')
evaluations = llm_evaluate_summaries(
chat_message.question, summaries, chat_message.model)
# 3. pull in the top documents from summaries
if evaluations:
response = commons['supabase'].from_('vectors').select(
'*').in_('id', values=[e['document_id'] for e in evaluations]).execute()
# 4. use top docs as additional context
additional_context = '---\nAdditional Context={}'.format(
'---\n'.join(data['content'] for data in response.data)
) + '\n'
model_response = qa(
{"question": additional_context + chat_message.question})
model_response = qa(
{"question": additional_context + chat_message.question})
else:
transformed_history = []
# Iterate through pairs in the history (assuming each user message is followed by an assistant message)
for i in range(0, len(chat_message.history) - 1, 2):
user_message = chat_message.history[i][1]
assistant_message = chat_message.history[i + 1][1]
transformed_history.append((user_message, assistant_message))
model_response = qa({"question": chat_message.question, "chat_history":transformed_history})
model_response = qa({"question": chat_message.question, "chat_history": transformed_history})
answer = model_response['answer']
# append sources (file_name) to answer
if "source_documents" in answer:
# logger.debug('Source Documents: %s', answer["source_documents"])
sources = [
doc.metadata["file_name"] for doc in answer["source_documents"]
if "file_name" in doc.metadata]
# logger.debug('Sources: %s', sources)
if sources:
files = dict.fromkeys(sources)
# # shall provide file links until pages available
# files = [f"[{f}](/explore/{f})" for f in files]
answer = answer + "\n\nRef: " + "; ".join(files)
return answer
return answer