import os
import streamlit as st
from typing_extensions import TypedDict, List
from IPython.display import Image, display
from langchain_core.pydantic_v1 import BaseModel, Field
from langchain.schema import Document
from langgraph.graph import START, END, StateGraph
from langchain.prompts import PromptTemplate
import uuid
from langchain_groq import ChatGroq
from langchain_community.utilities import GoogleSerperAPIWrapper
from langchain_chroma import Chroma
from langchain_community.document_loaders import NewsURLLoader
from langchain_community.retrievers.wikipedia import WikipediaRetriever
from sentence_transformers import SentenceTransformer
from langchain.vectorstores import Chroma
from langchain_community.document_loaders import UnstructuredURLLoader, NewsURLLoader
from langchain_community.embeddings import HuggingFaceEmbeddings
from langchain.text_splitter import RecursiveCharacterTextSplitter
from langchain_community.document_loaders import WebBaseLoader
from langchain_core.output_parsers import StrOutputParser
from langchain_core.output_parsers import JsonOutputParser
from langchain_community.vectorstores.utils import filter_complex_metadata
from langchain.schema import Document
from langchain_community.document_loaders.directory import DirectoryLoader
from langchain.document_loaders import TextLoader
from langgraph.graph import START, END, StateGraph
from langchain.retrievers import WebResearchRetriever
from langchain.callbacks.manager import CallbackManager
from langchain.callbacks.streaming_stdout import StreamingStdOutCallbackHandler
from exa_py import Exa
os.environ["LANGCHAIN_TRACING_V2"]="true"
os.environ["LANGCHAIN_ENDPOINT"]= "https://api.smith.langchain.com"
os.environ["LANGCHAIN_PROJECT"] = "Civilinės_teises_Asistente_V1_Embed"
lang_api_key = os.getenv("LANGCHAIN_API_KEY")
SERPER_API_KEY = os.getenv("SERPER_API_KEY")
groq_api_key = os.getenv("GROQ_API_KEY")
exa_api_key = os.getenv("exa_api_key")
exa = Exa(api_key="exa_api_key")
def create_retriever_from_chroma(vectorstore_path="docs/chroma/", search_type='mmr', k=7, chunk_size=300, chunk_overlap=30):
model_name = "Alibaba-NLP/gte-multilingual-base"
model_kwargs = {'device': 'cpu',
"trust_remote_code" : 'False'}
encode_kwargs = {'normalize_embeddings': True}
embeddings = HuggingFaceEmbeddings(
model_name=model_name,
model_kwargs=model_kwargs,
encode_kwargs=encode_kwargs
)
if os.path.exists(vectorstore_path) and os.listdir(vectorstore_path):
vectorstore = Chroma(persist_directory=vectorstore_path,embedding_function=embeddings)
else:
st.write("Vector store doesnt exist and will be created now")
loader = DirectoryLoader('./data/', glob="./*.txt", loader_cls=TextLoader)
docs = loader.load()
text_splitter = RecursiveCharacterTextSplitter.from_tiktoken_encoder(
chunk_size=chunk_size, chunk_overlap=chunk_overlap,
separators=["\n\n \n\n","\n\n\n", "\n\n", r"In \[[0-9]+\]", r"\n+", r"\s+"],
is_separator_regex = True
)
split_docs = text_splitter.split_documents(docs)
vectorstore = Chroma.from_documents(
documents=split_docs, embedding=embeddings, persist_directory=vectorstore_path,
)
retriever=vectorstore.as_retriever(search_type = search_type, search_kwargs={"k": k})
return retriever
def handle_userinput(user_question, custom_graph):
# Add the user's question to the chat history and display it in the UI
st.session_state.messages.append({"role": "user", "content": user_question})
st.chat_message("user").write(user_question)
# Generate a unique thread ID for the graph's state
config = {"configurable": {"thread_id": str(uuid.uuid4())}}
try:
# Invoke the custom graph with the input question
state_dict = custom_graph.invoke(
{"question": user_question, "steps": []}, config
)
docs = state_dict["documents"]
with st.sidebar:
st.subheader("Dokumentai, kuriuos Birutė gavo kaip kontekstą")
with st.spinner("Processing"):
for doc in docs:
# Extract document content
content = doc
# Extract document metadata if available
#metadata =doc.metadata.get('original_doc_name', 'unknown')
# Display content and metadata
st.write(f"Documentas: {content}")
# Check if a response (generation) was produced by the graph
if 'generation' in state_dict and state_dict['generation']:
response = state_dict["generation"]
# Add the assistant's response to the chat history and display it
st.session_state.messages.append({"role": "assistant", "content": response})
st.chat_message("assistant").write(response)
else:
st.chat_message("assistant").write("Your question violates toxicity rules or contains sensitive information.")
except Exception as e:
# Display an error message in case of failure
st.chat_message("assistant").write("Klaida: Arba per didelis kontekstas suteiktas modeliui, arba užklausų serveryje yra per daug")
from typing import Annotated
def create_workflow(retriever):
class GraphState(TypedDict):
"""
Represents the state of our graph.
Attributes:
question: question
generation: LLM generation
search: whether to add search
documents: list of documents
generations_count : generations count
"""
question: Annotated[str, "Single"] # Ensuring only one value per step
generation: str
search: str
documents: List[str]
steps: List[str]
generation_count: int
llm = ChatGroq(
model="llama-3.3-70b-versatile",
temperature=0.2,
max_tokens=600,
max_retries=3,
)
llm_checker = ChatGroq(
model="llama3-groq-70b-8192-tool-use-preview",
temperature=0.1,
max_tokens=400,
max_retries=3,
)
workflow = StateGraph(GraphState)
# Define the nodes
workflow.add_node("ask_question", lambda state: ask_question(state))
workflow.add_node("retrieve", lambda state: retrieve(state, retriever))
workflow.add_node("grade_documents", lambda state: grade_documents(state, retrieval_grader_grader(llm_checker)))
workflow.add_node("generate", lambda state: generate(state, QA_chain(llm)))
workflow.add_node("web_search", web_search)
#workflow.add_node("transform_query", lambda state: transform_query(state, create_question_rewriter(llm)))
# Build graph
workflow.set_entry_point("ask_question")
workflow.add_edge("ask_question", "retrieve")
workflow.add_edge("retrieve", "grade_documents")
#workflow.add_edge("retrieve", "generate")
workflow.add_conditional_edges(
"grade_documents",
decide_to_generate,
{
"search": "web_search",
"generate": "generate",
},
)
workflow.add_edge("web_search", "generate")
workflow.add_edge("generate", END)
custom_graph = workflow.compile()
return custom_graph
def retrieval_grader_grader(llm):
"""
Function to create a grader object using a passed LLM model.
Args:
llm: The language model to be used for grading.
Returns:
Callable: A pipeline function that grades relevance based on the LLM.
"""
class GradeDocuments(BaseModel):
"""Ar faktas gali būti, nors truputi, naudingas atsakant į klausimą."""
binary_score: str = Field(
description="Documentai yra aktualūs klausimui, 'yes' arba 'no'"
)
# Create the structured LLM grader using the passed LLM
structured_llm_grader = llm.with_structured_output(GradeDocuments)
# Define the prompt template
prompt = PromptTemplate(
template="""Jūs esate mokytojas, vertinantis viktoriną. Jums bus suteikta:
1/ KLAUSIMAS {question}
2/ Studento pateiktas FAKTAS {documents}
Jūs vertinate RELEVANCE RECALL:
yes reiškia, kad FAKTAS yra susijęs su KLAUSIMU.
no reiškia, kad FAKTAS nesusijęs su KLAUSIMU.
yes yra aukščiausias (geriausias) balas. no yra žemiausias balas, kurį galite duoti.
Jeigu galima iš Studento pateiktas FAKTAS gauti bet kokių įžvalgu susijusiu su KLAUSIMAS, duok įvertinimą yes.
Žingsnis po žingsnio paaiškinkite savo samprotavimus. Įsitikinkite, kad jūsų samprotavimai ir išvados yra teisingi.
Iš pradžių venkite tiesiog nurodyti teisingą atsakymą.
Klausimas: {question} \n
FAKTAS: \n\n {documents} \n\n
Suteikite dvejetainį balą „yes“ arba „no“, kad nurodytumėte, ar dokumentas yra susijęs su klausimu. \n
Pateikite dvejetainį balą kaip JSON su vienu raktu „balu“ ir be įžangos ar paaiškinimo.
""",
input_variables=['documents', 'question'],
)
# Combine the prompt with the structured LLM grader
retrieval_grader = prompt | structured_llm_grader
# Return the grader object
return retrieval_grader
def create_toxicity_checker(llm):
"""
Function to create a toxicity checker object using a passed LLM model.
Args:
llm: The language model to be used for toxicity checking.
Returns:
Callable: A pipeline function that checks for toxicity based on the LLM.
"""
# Define the class for toxicity check inside the function
class PromtGuard(BaseModel):
"""Binary score for toxicity check on question."""
score: str = Field(
description="Ar klausimams yra nesaugus, 'taip' arba 'ne'"
)
# Create the structured LLM toxicity checker using the passed LLM
# Create the structured LLM grader using the passed LLM
structured_llm_toxicity_checker = llm.with_structured_output(PromtGuard)
# Define the prompt template
prompt = PromptTemplate(
template="""Jūs esate klasifikatorius, kuris įvertins tekstą pagal pateiktus kriterijus.
Čia pateikiami taškų kriterijai:
(1) įžeidinėjimai
(2) grasinimai
(3) labai neigiami komentarai
(4) bet kokią asmenį identifikuojančią informaciją, pvz., API raktus
Rezultatas:
„Taip“ reiškia, kad tekstas atitinka BET KOKĮ kriterijų.
Įvertinimas „ne“ reiškia, kad tekstas neatitinka VISŲ kriterijų.
Štai klausimas: {question}
Suteikite dvejetainį balą „taip“ arba „ne“, kad nurodytumėte, ar atsakymas yra naudingas norint išspręsti klausimą.
Jei atsakyme yra pasikartojančių frazių, kartojimas, tada grąžinkite „ne“\n
Pateikite dvejetainį balą kaip JSON su vienu raktu „balu“ ir be įžangos ar paaiškinimo.""",
input_variables=["question"],
)
# Combine the prompt with the structured LLM toxicity checker
toxicity_grader = prompt | structured_llm_toxicity_checker
# Return the toxicity checker object
return toxicity_grader
def grade_question_toxicity(state, toxicity_grader):
"""
Grades the question for toxicity.
Args:
state (dict): The current graph state.
Returns:
str: 'good' if the question passes the toxicity check, 'bad' otherwise.
"""
steps = state["steps"]
steps.append("promt guard")
score = toxicity_grader.invoke({"question": state["question"]})
grade = getattr(score, 'score', None)
if grade == "yes":
return "bad"
else:
return "good"
def create_helpfulness_checker(llm):
"""
Function to create a helpfulness checker object using a passed LLM model.
Args:
llm: The language model to be used for checking the helpfulness of answers.
Returns:
Callable: A pipeline function that checks if the student's answer is helpful.
"""
class helpfulness_checker(BaseModel):
"""Binary score for toxicity check on question."""
score: str = Field(
description="Ar atsakymas yra naudingas?, 'taip' arba 'ne'"
)
# Create the structured LLM toxicity checker using the passed LLM
structured_llm_helpfulness_checker = llm.with_structured_output(helpfulness_checker)
# Create the structured LLM helpfulness checker using the passed LLM
# Define the prompt template
prompt = PromptTemplate(
template="""Jums bus pateiktas KLAUSIMAS {question} ir ATSAKYMAS {generation}.
Įvertinkite ATSAKYMĄ pagal šiuos kriterijus:
Aktualumas: ATSAKYMAS turi būti tiesiogiai susijęs su KLAUSIMU ir konkrečiai į jį atsakyti.
Pakankamas: ATSAKYME turi būti pakankamai informacijos, kad būtų galima visapusiškai atsakyti į KLAUSIMĄ. Jei ATSAKYME vartojamos tokios frazės kaip „nežinau“, „neturiu pakankamai informacijos“, „pateiktuose dokumentuose apie tai neužsimenama“ ar panašių posakių, kuriuose vengiama tiesiogiai atsakyti į KLAUSIMĄ, įvertinkite „ne“.
Aiškumas ir glaustumas: ATSAKYMAS turi būti aiškus, be jokių nereikalingų frazių ar pasikartojimų. Jei jame yra perteklinė arba netiesioginė informacija, o ne tiesioginis atsakymas, įvertinkite „ne“.
Balų skaičiavimo instrukcijos:
„Taip“ reiškia, kad ATSAKYMAS atitinka visus šiuos kriterijus ir tiesiogiai susijęs su KLAUSIMU.
Įvertinimas „ne“ reiškia, kad ATSAKYMAS neatitinka visų šių kriterijų.
Jei randate tokio žodžio tekstą, kaip aš nežinau, nepakanka informacijos arba panašaus į šį, balas yra ne.
Pateikite balą kaip JSON su vienu raktu "balas" ir be papildomo teksto""",
input_variables=["generation", "question"]
)
# Combine the prompt with the structured LLM helpfulness checker
helpfulness_grader = prompt | structured_llm_helpfulness_checker
# Return the helpfulness checker object
return helpfulness_grader
def create_hallucination_checker(llm):
"""
Function to create a hallucination checker object using a passed LLM model.
Args:
llm: The language model to be used for checking hallucinations in the student's answer.
Returns:
Callable: A pipeline function that checks if the student's answer contains hallucinations.
"""
class hallucination_checker(BaseModel):
"""Binary score for toxicity check on question."""
score: str = Field(
description="Ar dokumentas yra susijes su atsakymu?, 'taip' arba 'ne'"
)
# Create the structured LLM toxicity checker using the passed LLM
structured_llm_hallucination_checker = llm.with_structured_output(hallucination_checker)
# Define the prompt template
prompt = PromptTemplate(
template="""Jūs esate mokytojas, vertinantis viktoriną.
Jums bus pateikti FAKTAI ir MOKINIO ATSAKYMAS.
Jūs vertinate MOKINIO ATSAKYMĄ iš šaltinio FAKTAI. Sutelkite dėmesį į MOKINIO ATSAKYMO teisingumą ir bet kokių haliucinacijų aptikimą.
Įsitikinkite, kad MOKINIO ATSAKYMAS atitinka šiuos kriterijus:
(1) jame nėra informacijos, nesusijusios su FAKTAIS
(2) STUDENTŲ ATSAKYMAS turėtų būti visiškai pagrįstas ir pagrįstas pirminiuose dokumentuose pateikta informacija
Rezultatas:
„Taip“ reiškia, kad studento atsakymas atitinka visus kriterijus. Tai aukščiausias (geriausias) balas.
Balas „ne“ reiškia, kad studento atsakymas neatitinka visų kriterijų. Tai yra žemiausias galimas balas, kurį galite duoti.
Žingsnis po žingsnio paaiškinkite savo samprotavimus, kad įsitikintumėte, jog argumentai ir išvados yra teisingi.
Iš pradžių venkite tiesiog nurodyti teisingą atsakymą.
MOKINIO ATSAKYMAS: {generation} \n
FAKTAI: \n\n {documents} \n\n
Suteikite dvejetainį balą „taip“ arba „ne“, kad nurodytumėte, ar dokumentas yra susijęs su klausimu. \n
Pateikite dvejetainį balą kaip JSON su vienu raktu „balu“ ir be įžangos ar paaiškinimo.""",
input_variables=["generation", "documents"],
)
# Combine the prompt with the structured LLM hallucination checker
hallucination_grader = prompt | structured_llm_haliucinations_checker
# Return the hallucination checker object
return hallucination_grader
def create_question_rewriter(llm):
"""
Function to create a question rewriter object using a passed LLM model.
Args:
llm: The language model to be used for rewriting questions.
Returns:
Callable: A pipeline function that rewrites questions for optimized vector store retrieval.
"""
# Define the prompt template for question rewriting
re_write_prompt = PromptTemplate(
template="""Esate klausimų perrašytojas, kurio specializacija yra Lietuvos teisė, tobulinanti klausimus, kad būtų galima optimizuoti jų paiešką iš teisinių dokumentų. Jūsų tikslas – išaiškinti teisinę intenciją, pašalinti dviprasmiškumą ir pakoreguoti formuluotes taip, kad jos atspindėtų teisinę kalbą, daugiausia dėmesio skiriant atitinkamiems raktiniams žodžiams, siekiant užtikrinti tikslų informacijos gavimą iš Lietuvos teisės šaltinių.
Man nereikia paaiškinimų, tik perrašyto klausimo.
Štai pradinis klausimas: \n\n {question}. Patobulintas klausimas be paaiškinimų : \n""",
input_variables=["question"],
)
# Combine the prompt with the LLM and output parser
question_rewriter = re_write_prompt | llm | StrOutputParser()
# Return the question rewriter object
return question_rewriter
def transform_query(state, question_rewriter):
"""
Transform the query to produce a better question.
Args:
state (dict): The current graph state
Returns:
state (dict): Updates question key with a re-phrased question
"""
print("---TRANSFORM QUERY---")
question = state["question"]
documents = state["documents"]
steps = state["steps"]
steps.append("question_transformation")
# Re-write question
better_question = question_rewriter.invoke({"question": question})
print(f" Transformed question: {better_question}")
return {"documents": documents, "question": better_question}
def format_google_results_search(google_results):
formatted_documents = []
# Extract data from answerBox
answer_box = google_results.get("answerBox", {})
answer_box_title = answer_box.get("title", "No title")
answer_box_answer = answer_box.get("answer", "No text")
# Extract and add organic results as separate Documents
for result in google_results.get("organic", []):
title = result.get("title", "No title")
link = result.get("link", "Nėra svetainės adreso")
snippet = result.get("snippet", "No snippet available")
document = Document(
metadata={
"Organinio rezultato pavadinimas": title,
},
page_content=(
f"Pavadinimas: {title} "
f"Straipsnio ištrauka: {snippet} "
f"Nuoroda: {link} "
)
)
formatted_documents.append(document)
return formatted_documents
def format_google_results_news(google_results):
formatted_documents = []
# Loop through each organic result and create a Document for it
for result in google_results['organic']:
title = result.get('title', 'No title')
link = result.get('link', 'No link')
descripsion = result.get('description', 'No link')
snippet = result.get('snippet', 'No summary available')
text = result.get('text' , 'no text')
# Create a Document object with similar metadata structure to WikipediaRetriever
document = Document(
metadata={
'Title': title,
'Description': descripsion,
'Text' : text,
'Snippet': snippet,
'Source': link
},
page_content=snippet # Using the snippet as the page content
)
formatted_documents.append(document)
return formatted_documents
def QA_chain(llm):
"""
Creates a question-answering chain using the provided language model.
Args:
llm: The language model to use for generating answers.
Returns:
An LLMChain configured with the question-answering prompt and the provided model.
"""
# Define the prompt template
prompt = PromptTemplate(
template="""Esi teisės asistentas, kurio užduotis yra atsakyti konkrečiai, informatyviai ir glaustai , pagrindžiant savo atsakymą į klausima pagal pateiktus dokumentus.
Atsakymas turi būti lietuvių kalba. Nesikartok.
Jei negali atsakyti į klausimą, pasakyk, Atsiprašau, nežinau atsakymo į jūsų klausimą.
Neužduok papildomų klausimų.
Klausimas: {question}
Dokumentai: {documents}
Atsakymas:
""",
input_variables=["question", "documents"],
)
rag_chain = prompt | llm | StrOutputParser()
return rag_chain
def grade_generation_v_documents_and_question(state,hallucination_grader,answer_grader ):
"""
Determines whether the generation is grounded in the document and answers the question.
"""
print("---CHECK HALLUCINATIONS---")
question = state["question"]
documents = state["documents"]
generation = state["generation"]
generation_count = state.get("generation_count") # Use state.get to avoid KeyError
print(f" generation number: {generation_count}")
# Grading hallucinations
score = hallucination_grader.invoke(
{"documents": documents, "generation": generation}
)
grade = getattr(score, 'score', None)
# Check hallucination
if grade == "yes":
print("---DECISION: GENERATION IS GROUNDED IN DOCUMENTS---")
# Check question-answering
print("---GRADE GENERATION vs QUESTION---")
score = answer_grader.invoke({"question": question, "generation": generation})
grade = getattr(score, 'score', None)
if grade == "yes":
print("---DECISION: GENERATION ADDRESSES QUESTION---")
return "useful"
else:
print("---DECISION: GENERATION DOES NOT ADDRESS QUESTION---")
return "not useful"
else:
if generation_count > 1:
print("---DECISION: GENERATION IS NOT GROUNDED IN DOCUMENTS, TRANSFORM QUERY---")
# Reset count if it exceeds limit
return "not useful"
else:
print("---DECISION: GENERATION IS NOT GROUNDED IN DOCUMENTS, RE-TRY---")
# Increment correctly here
print(f" generation number after increment: {state['generation_count']}")
return "not supported"
def ask_question(state):
"""
Initialize question
Args:
state (dict): The current graph state
Returns:
state (dict): Question
"""
steps = state["steps"]
question = state["question"]
generations_count = state.get("generations_count", 0)
steps.append("question_asked")
return {"question": question, "steps": steps,"generation_count": generations_count}
def retrieve(state , retriever):
"""
Retrieve documents
Args:
state (dict): The current graph state
retriever: The retriever object
Returns:
state (dict): New key added to state, documents, that contains retrieved documents
"""
steps = state["steps"]
question = state["question"]
documents = retriever.invoke(question)
steps.append("retrieve_documents")
return {"documents": documents, "question": question, "steps": steps}
def generate(state,QA_chain):
"""
Generate answer
"""
question = state["question"]
documents = state["documents"]
generation = QA_chain.stream({"documents": documents, "question": question})
steps = state["steps"]
steps.append("generate_answer")
generation_count = state["generation_count"]
generation_count += 1
return {
"documents": documents,
"question": question,
"generation": generation,
"steps": steps,
"generation_count": generation_count # Include generation_count in return
}
def grade_documents(state, retrieval_grader):
question = state["question"]
documents = state["documents"]
steps = state["steps"]
steps.append("grade_document_retrieval")
filtered_docs = []
web_results_list = []
search = "No"
for d in documents:
# Call the grading function
score = retrieval_grader.invoke({"question": question, "documents": d})
print(f"Grader output for document: {score}") # Detailed debugging output
# Extract the grade
grade = getattr(score, 'binary_score', None)
if grade and grade.lower() in ["yes", "true", "1",'taip']:
filtered_docs.append(d)
elif len(filtered_docs) < 4:
search = "Yes"
# Check the decision-making process
print(f"Final decision - Perform web search: {search}")
print(f"Filtered documents count: {len(filtered_docs)}")
return {
"documents": filtered_docs,
"question": question,
"search": search,
"steps": steps,
}
def clean_exa_document(doc):
"""
Extracts and retains only the title, url, text, and summary from the exa result document.
"""
return {
" Pavadinimas: ": doc.title,
" Apibendrinimas: ": doc.summary,
" Straipnsio internetinis adresas: ": doc.url,
" Tekstas: ": doc.text
}
def web_search(state):
question = state["question"]
documents = state.get("documents", [])
steps = state["steps"]
steps.append("web_search")
k = 8 - len(documents)
web_results_list = []
# Fetch results from exa
exa_results_raw = exa.search_and_contents(
query=question,
start_published_date="2018-01-01T22:00:01.000Z",
type="keyword",
num_results=2,
text={"max_characters": 7000},
summary={
"query": "Tell in summary a meaning about what is article written. This summary has to be written in a way to be related to {question} Provide facts, be concise. Do it in Lithuanian language."
},
include_domains=[ "infolex.lt", "vmi.lt", "lrs.lt", "e-seimas.lrs.lt", "teise.pro",'lt.wikipedia.org', 'teismai.lt' ],
)
# Extract results
exa_results = exa_results_raw.results if hasattr(exa_results_raw, "results") else []
cleaned_exa_results = [clean_exa_document(doc) for doc in exa_results]
if len(cleaned_exa_results) <1:
web_results = GoogleSerperAPIWrapper(k=2, gl="lt", hl="lt", type="search").results(question)
formatted_documents = format_google_results_search(web_results)
web_results_list.extend(formatted_documents if isinstance(formatted_documents, list) else [formatted_documents])
combined_documents = documents + cleaned_exa_results +web_results_list
else:
combined_documents = documents + cleaned_exa_results
return {"documents": combined_documents, "question": question, "steps": steps}
def decide_to_generate(state):
"""
Determines whether to generate an answer, or re-generate a question.
Args:
state (dict): The current graph state
Returns:
str: Binary decision for next node to call
"""
search = state["search"]
if search == "Yes":
return "search"
else:
return "generate"
def decide_to_generate2(state):
"""
Determines whether to generate an answer, or re-generate a question.
Args:
state (dict): The current graph state
Returns:
str: Binary decision for next node to call
"""
search = state["search"]
if search == "Yes":
return "search"
else:
return "generate"