marigen_api / api.py
jameszokah's picture
Synced repo using 'sync_with_huggingface' Github Action
84bd7df verified
import datetime
import os
from dotenv import load_dotenv
import asyncio
from fastapi import FastAPI, Body, File, UploadFile, HTTPException
from fastapi.responses import StreamingResponse
from typing import List, AsyncIterable, Annotated, Optional
from enum import Enum
from fastapi.middleware.cors import CORSMiddleware
from pydantic import BaseModel
from langchain_openai import ChatOpenAI
from langchain import hub
from langchain_chroma import Chroma
from langchain_core.output_parsers import StrOutputParser
from langchain_core.runnables import RunnablePassthrough
from langchain_nomic.embeddings import NomicEmbeddings
from langchain_text_splitters import RecursiveCharacterTextSplitter
from langchain.callbacks import AsyncIteratorCallbackHandler
from langchain_core.documents import Document
from in_memory import load_all_documents
from langchain_nomic.embeddings import Embeddings, NomicEmbeddings
from loader import load_web_content, load_youtube_content
from get_pattern import generate_pattern
from get_agents import process_agents
# ################################### FastAPI setup ############################################
app = FastAPI()
origins = ["*"]
app.add_middleware(
CORSMiddleware,
allow_origins=origins,
allow_credentials=True,
allow_methods=["*"],
allow_headers=["*"],
)
# ################################### Helper functions ############################################
# async def load_all_documents(files: List[UploadFile]) -> List[Document]:
# documents = []
# for file in files:
# docs = await load_document(file)
# documents.extend(docs)
# return documents
# ################################### LLM, RAG and Streaming ############################################
load_dotenv()
GROQ_API_KEY = os.environ.get("GROQ_API_KEY")
GROQ_API_BASE = os.environ.get("GROQ_API_BASE")
OPENAI_MODEL_NAME = os.environ.get("OPENAI_MODEL_NAME")
embedding_model = NomicEmbeddings(model="nomic-embed-text-v1.5")
def split_documents(documents: List[Document], chunk_size=1000, chunk_overlap=200) -> List[Document]:
text_splitter = RecursiveCharacterTextSplitter(
chunk_size=chunk_size, chunk_overlap=chunk_overlap
)
print("Splitting documents into chunks...")
return text_splitter.split_documents(documents)
def generate_embeddings(documents: List[Document]) -> NomicEmbeddings:
embedding_model = NomicEmbeddings(model="nomic-embed-text-v1.5")
embeddings = [embedding_model.embed(
[document.page_content], task_type='search_document') for document in documents]
return embedding_model
def store_embeddings(documents: List[Document], embeddings: NomicEmbeddings):
vectorstore = Chroma.from_documents(
documents=documents, embedding=embeddings, persist_directory="./chroma_db")
return vectorstore
def load_embeddings(embeddings: NomicEmbeddings) -> Chroma:
embeddings = Chroma(persist_directory="./chroma_db",
embedding_function=embeddings)
return embeddings
# ################################### Updated generate_chunks Function ############################################
async def generate_chunks(query: str) -> AsyncIterable[str]:
callback = AsyncIteratorCallbackHandler()
llm = ChatOpenAI(
openai_api_base=GROQ_API_BASE,
api_key=GROQ_API_KEY,
temperature=0.0,
model_name=OPENAI_MODEL_NAME, # "mixtral-8x7b-32768",
streaming=True, # ! important
verbose=True,
callbacks=[callback]
)
# Load vector store (this should be pre-populated with documents and embeddings)
# Ensure to modify this to load your actual vector store
vectorstore = load_embeddings(embeddings=embedding_model)
# Retrieve relevant documents for the query
retriever = vectorstore.as_retriever()
# relevant_docs = retriever(query)
# Combine the retrieved documents into a single string
def format_docs(docs):
return "\n\n".join(doc.page_content for doc in docs)
# Define the RAG chain
prompt = hub.pull("rlm/rag-prompt")
rag_chain = (
{"context": retriever | format_docs, "question": RunnablePassthrough()}
| prompt
| llm
| StrOutputParser()
)
# Generate the response
task = asyncio.create_task(
rag_chain.ainvoke(query)
)
index = 0
try:
async for token in callback.aiter():
print(index, ": ", token, ": ", datetime.datetime.now().time())
index = index + 1
yield token
except Exception as e:
print(f"Caught exception: {e}")
finally:
callback.done.set()
await task
# ################################### Models ########################################
class QuestionType(str, Enum):
PATTERN = "PATTERN"
AGENTS = "AGENTS"
RAG = "RAG"
class Input(BaseModel):
question: str
type: QuestionType
pattern: Optional[str]
chat_history: List[str]
class Metadata(BaseModel):
conversation_id: str
class Config(BaseModel):
metadata: Metadata
class RequestBody(BaseModel):
input: Input
config: Config
# ################################### Routes ############################################
@app.get("/")
def read_root():
return {"Hello": "World from Marigen"}
@app.post("/chat", response_class=StreamingResponse)
async def chat(query: RequestBody = Body(...)):
print(query.input.question)
print(query.input.type)
if query.input.type == QuestionType.PATTERN:
print(query.input.pattern)
pattern = query.input.pattern
gen = generate_pattern(pattern=pattern, query=query.input.question)
return StreamingResponse(gen, media_type="text/event-stream")
elif query.input.type == QuestionType.AGENTS:
gen = process_agents(query.input.question)
return StreamingResponse(gen, media_type="text/event-stream")
elif query.input.type == QuestionType.RAG:
gen = generate_chunks(query.input.question)
return StreamingResponse(gen, media_type="text/event-stream")
raise HTTPException(status_code=400, detail="No accurate response for your given query")
@app.post("/uploadfiles")
async def create_upload_files(
files: Annotated[List[UploadFile], File(description="Multiple files as UploadFile")],
):
try:
# Load documents from files
documents = await load_all_documents(files)
print(f"Loaded {len(documents)} documents")
print(f"----------> {documents} documents <-----------")
chunks = []
# Split documents into chunks
for docs in documents:
print(docs)
chunk = split_documents(docs[0])
chunks.extend(chunk)
print(f"Split into {len(chunks)} chunks")
# Generate embeddings for chunks
# embeddings_model = generate_embeddings(chunks)
# print(f"Generated {len(embeddings)} embeddings")
# # Store embeddings in vector store
vectorstore = store_embeddings(chunks, embedding_model)
print("Embeddings stored in vector store")
return {"filenames": [file.filename for file in files], 'chunks': chunks, "message": "Files processed and embeddings generated."}
except Exception as e:
print(f"Error loading documents: {e}")
return {"message": f"Error loading documents: {e}"}
# New routes for YouTube and website content loading
@app.post("/load_youtube")
async def load_youtube(youtube_url: str):
try:
documents = load_youtube_content(youtube_url)
chunks = split_documents(documents)
store_embeddings(chunks, embedding_model)
return {"message": f"YouTube video loaded and processed successfully.", "documents": documents}
except Exception as e:
print(f"Error loading YouTube video: {e}")
return {"message": f"Error loading YouTube video: {e}"}
@app.post("/load_website")
async def load_website(website_url: str):
try:
documents = load_web_content(website_url)
chunks = split_documents(documents)
store_embeddings(chunks, embedding_model)
return {"message": f"Website loaded and processed successfully.", "documents": documents}
except Exception as e:
print(f"Error loading website: {e}")
return {"message": f"Error loading website: {e}"}
@app.post("/query")
async def query_vector_store(query: str):
# Load the vector store (ensure you maintain a reference to it, possibly store in memory or a persistent store)
# Modify this with actual loading mechanism
vectorstore = load_embeddings(embeddings=embedding_model)
# Perform a query to retrieve relevant documents
relevant_docs = vectorstore.query(query)
return {"query": query, "results": [doc.page_content for doc in relevant_docs]}