pathfinder / app_gradio.py
kiyer's picture
added option for reranking without HyDE
b95eb79 verified
import gradio as gr
import numpy as np
from abc import ABC, abstractmethod
from typing import List, Dict, Any, Tuple
from collections import defaultdict
import pandas as pd
from datetime import datetime, date
from datasets import load_dataset, load_from_disk
from collections import Counter
import yaml, json, requests, sys, os, time
import urllib.parse
import concurrent.futures
from langchain import hub
from langchain_openai import ChatOpenAI as openai_llm
from langchain_openai import OpenAIEmbeddings
from langchain_core.runnables import RunnableConfig, RunnablePassthrough, RunnableParallel
from langchain_core.prompts import PromptTemplate
from langchain_community.callbacks import StreamlitCallbackHandler
from langchain_community.utilities import DuckDuckGoSearchAPIWrapper
from langchain_community.vectorstores import Chroma
from langchain_community.document_loaders import TextLoader
from langchain.agents import create_react_agent, Tool, AgentExecutor
from langchain.text_splitter import RecursiveCharacterTextSplitter
from langchain_core.output_parsers import StrOutputParser
from langchain.callbacks import FileCallbackHandler
from langchain.callbacks.manager import CallbackManager
from langchain.schema import Document
import instructor
from pydantic import BaseModel, Field
from typing import List, Literal
from nltk.corpus import stopwords
import nltk
from openai import OpenAI, moderations
# import anthropic
import cohere
import faiss
import matplotlib.pyplot as plt
import spacy
from string import punctuation
import pytextrank
from prompts import *
openai_key = os.environ['openai_key']
cohere_key = os.environ['cohere_key']
os.environ["OPENAI_API_KEY"] = os.environ['openai_key']
def load_nlp():
nlp = spacy.load("en_core_web_sm")
nlp.add_pipe("textrank")
try:
stopwords.words('english')
except:
nltk.download('stopwords')
stopwords.words('english')
return nlp
gen_llm = openai_llm(temperature=0, model_name='gpt-4o-mini', openai_api_key = openai_key)
consensus_client = instructor.patch(OpenAI(api_key=openai_key))
embed_client = OpenAI(api_key = openai_key)
embed_model = "text-embedding-3-small"
embeddings = OpenAIEmbeddings(model = embed_model, api_key = openai_key)
nlp = load_nlp()
def check_mod(query):
mod_report = moderations.create(input=query)
for i in mod_report.results[0].categories:
if i[1] == True:
return True
return False
def get_keywords(text, nlp=nlp):
result = []
pos_tag = ['PROPN', 'ADJ', 'NOUN']
doc = nlp(text.lower())
for token in doc:
if(token.text in nlp.Defaults.stop_words or token.text in punctuation):
continue
if(token.pos_ in pos_tag):
result.append(token.text)
return result
def load_arxiv_corpus():
# arxiv_corpus = load_from_disk('data/')
# arxiv_corpus.load_faiss_index('embed', 'data/astrophindex.faiss')
# keeping it up to date with the dataset
arxiv_corpus = load_dataset('kiyer/pathfinder_arxiv_data', split='train')
arxiv_corpus.add_faiss_index(column='embed')
print('loading arxiv corpus from disk')
return arxiv_corpus
class RetrievalSystem():
def __init__(self):
self.dataset = arxiv_corpus
self.client = OpenAI(api_key = openai_key)
self.embed_model = "text-embedding-3-small"
self.generation_client = openai_llm(temperature=0,model_name='gpt-4o-mini', openai_api_key = openai_key)
self.hyde_client = openai_llm(temperature=0.5,model_name='gpt-4o-mini', openai_api_key = openai_key)
self.cohere_client = cohere.Client(cohere_key)
def make_embedding(self, text):
str_embed = self.client.embeddings.create(input = [text], model = self.embed_model).data[0].embedding
return str_embed
def embed_batch(self, texts: List[str]) -> List[np.ndarray]:
embeddings = self.client.embeddings.create(input=texts, model=self.embed_model).data
return [np.array(embedding.embedding, dtype=np.float32) for embedding in embeddings]
def get_query_embedding(self, query):
return self.make_embedding(query)
def calc_faiss(self, query_embedding, top_k = 100):
# xq = query_embedding.reshape(-1,1).T.astype('float32')
# D, I = self.index.search(xq, top_k)
# return I[0], D[0]
tmp = self.dataset.search('embed', query_embedding, k=top_k)
return [tmp.indices, tmp.scores, self.dataset[tmp.indices]]
def rank_and_filter(self, query, query_embedding, top_k = 10, top_k_internal = 1000, return_scores=False):
if 'Keywords' in self.toggles:
self.weight_keywords = True
else:
self.weight_keywords = False
if 'Time' in self.toggles:
self.weight_date = True
else:
self.weight_date = False
if 'Citations' in self.toggles:
self.weight_citation = True
else:
self.weight_citation = False
topk_indices, similarities, small_corpus = self.calc_faiss(np.array(query_embedding), top_k = top_k_internal)
similarities = 1/similarities # converting from a distance (less is better) to a similarity (more is better)
if self.weight_keywords == True:
query_kws = get_keywords(query)
input_kws = self.query_input_keywords
query_kws = query_kws + input_kws
self.query_kws = query_kws
sub_kws = [small_corpus['keywords'][i] for i in range(top_k_internal)]
kw_weight = np.zeros((len(topk_indices),)) + 0.1
for k in query_kws:
for i in (range(len(topk_indices))):
for j in range(len(sub_kws[i])):
if k.lower() in sub_kws[i][j].lower():
kw_weight[i] = kw_weight[i] + 0.1
# print(i, k, sub_kws[i][j])
# kw_weight = kw_weight**0.36 / np.amax(kw_weight**0.36)
kw_weight = kw_weight / np.amax(kw_weight)
else:
kw_weight = np.ones((len(topk_indices),))
if self.weight_date == True:
sub_dates = [small_corpus['date'][i] for i in range(top_k_internal)]
date = datetime.now().date()
date_diff = np.array([((date - i).days / 365.) for i in sub_dates])
# age_weight = (1 + np.exp(date_diff/2.1))**(-1) + 0.5
age_weight = (1 + np.exp(date_diff/0.7))**(-1)
age_weight = age_weight / np.amax(age_weight)
else:
age_weight = np.ones((len(topk_indices),))
if self.weight_citation == True:
# st.write('weighting by citations')
sub_cites = np.array([small_corpus['cites'][i] for i in range(top_k_internal)])
temp = sub_cites.copy()
temp[sub_cites > 300] = 300.
cite_weight = (1 + np.exp((300-temp)/42.0))**(-1.)
cite_weight = cite_weight / np.amax(cite_weight)
else:
cite_weight = np.ones((len(topk_indices),))
similarities = similarities * (kw_weight) * (age_weight) * (cite_weight)
filtered_results = [[topk_indices[i], similarities[i]] for i in range(len(similarities))]
top_results = sorted(filtered_results, key=lambda x: x[1], reverse=True)[:top_k]
top_scores = [doc[1] for doc in top_results]
top_indices = [doc[0] for doc in top_results]
small_df = self.dataset[top_indices]
if return_scores:
return {doc[0]: doc[1] for doc in top_results}, small_df
# Only keep the document IDs
top_results = [doc[0] for doc in top_results]
return top_results, small_df
def generate_doc(self, query: str):
prompt = """You are an expert astronomer. Given a scientific query, generate the abstract of an expert-level research paper
that answers the question. Stick to a maximum length of {} tokens and return just the text of the abstract and conclusion.
Do not include labels for any section. Use research-specific jargon.""".format(self.max_doclen)
messages = [("system",prompt,),("human", query),]
return self.hyde_client.invoke(messages).content
def generate_docs(self, query: str):
docs = []
for i in range(self.generate_n):
docs.append(self.generate_doc(query))
return docs
def embed_docs(self, docs: List[str]):
return self.embed_batch(docs)
def retrieve(self, query, top_k, return_scores = False,
embed_query=True, max_doclen=250,
generate_n=1, temperature=0.5,
rerank_top_k = 250):
if max_doclen * generate_n > 8191:
raise ValueError("Too many tokens. Please reduce max_doclen or generate_n.")
query_embedding = self.get_query_embedding(query)
if self.hyde == True:
self.max_doclen = max_doclen
self.generate_n = generate_n
self.hyde_client.temperature = temperature
self.embed_query = embed_query
docs = self.generate_docs(query)
# st.expander('Abstract generated with hyde', expanded=False).write(docs)
doc_embeddings = self.embed_docs(docs)
if self.embed_query:
query_emb = self.embed_docs([query])[0]
doc_embeddings.append(query_emb)
query_embedding = np.mean(np.array(doc_embeddings), axis = 0)
if self.rerank == True:
top_results, small_df = self.rank_and_filter(query,
query_embedding,
rerank_top_k,
return_scores = False)
# try:
docs_for_rerank = [small_df['abstract'][i] for i in range(rerank_top_k)]
if len(docs_for_rerank) == 0:
return []
reranked_results = self.cohere_client.rerank(
query=query,
documents=docs_for_rerank,
model='rerank-english-v3.0',
top_n=top_k
)
final_results = []
for result in reranked_results.results:
doc_id = top_results[result.index]
doc_text = docs_for_rerank[result.index]
score = float(result.relevance_score)
final_results.append([doc_id, "", score])
final_indices = [doc[0] for doc in final_results]
if return_scores:
return {result[0]: result[2] for result in final_results}, self.dataset[final_indices]
return [doc[0] for doc in final_results], self.dataset[final_indices]
# except:
# print('heavy load, please wait 10s and try again.')
else:
top_results, small_df = self.rank_and_filter(query,
query_embedding,
top_k,
return_scores = return_scores)
return top_results, small_df
def return_formatted_df(self, top_results, small_df):
df = pd.DataFrame(small_df)
df = df.drop(columns=['umap_x','umap_y','cite_bibcodes','ref_bibcodes'])
links = ['['+i+'](https://ui.adsabs.harvard.edu/abs/'+i+'/abstract)' for i in small_df['bibcode']]
# st.write(top_results[0:10])
scores = [top_results[i] for i in top_results]
indices = [i for i in top_results]
df.insert(1,'ADS Link',links,True)
df.insert(2,'Relevance',scores,True)
df.insert(3,'indices',indices,True)
df = df[['ADS Link','Relevance','date','cites','title','authors','abstract','keywords','ads_id','indices','embed']]
df.index += 1
return df
arxiv_corpus = load_arxiv_corpus()
ec = RetrievalSystem()
print('loaded retrieval system')
def Library(papers_df):
op_docs = ''
for i in range(len(papers_df)):
op_docs = op_docs + 'Paper %.0f:' %(i+1) + papers_df['title'][i+1] + '\n' + papers_df['abstract'][i+1] + '\n\n'
return op_docs
def run_rag_qa(query, papers_df, question_type):
loaders = []
documents = []
for i, row in papers_df.iterrows():
content = f"Paper {i+1}: {row['title']}\n{row['abstract']}\n\n"
metadata = {"source": row['ads_id']}
doc = Document(page_content=content, metadata=metadata)
documents.append(doc)
text_splitter = RecursiveCharacterTextSplitter(chunk_size=150, chunk_overlap=50, add_start_index=True)
splits = text_splitter.split_documents(documents)
vectorstore = Chroma.from_documents(documents=splits, embedding=embeddings, collection_name='retdoc4')
retriever = vectorstore.as_retriever(search_type="similarity", search_kwargs={"k": 6})
if question_type == 'Bibliometric':
template = bibliometric_prompt
elif question_type == 'Single-paper':
template = single_paper_prompt
elif question_type == 'Broad but nuanced':
template = deep_knowledge_prompt
else:
template = regular_prompt
prompt = PromptTemplate.from_template(template)
def format_docs(docs):
return "\n\n".join(doc.page_content for doc in docs)
rag_chain_from_docs = (
RunnablePassthrough.assign(context=(lambda x: format_docs(x["context"])))
| prompt
| gen_llm
| StrOutputParser()
)
rag_chain_with_source = RunnableParallel(
{"context": retriever, "question": RunnablePassthrough()}
).assign(answer=rag_chain_from_docs)
rag_answer = rag_chain_with_source.invoke(query, )
vectorstore.delete_collection()
# except:
# st.subheader('heavy load! please wait 10 seconds and try again.')
return rag_answer
def guess_question_type(query: str):
gen_client = openai_llm(temperature=0,model_name='gpt-4o-mini', openai_api_key = openai_key)
messages = [("system",question_categorization_prompt,),("human", query),]
return gen_client.invoke(messages).content
def log_to_gist(strings):
# Adding query logs to prevent and account for possible malicious use.
# Logs will be deleted periodically if not needed.
github_token = os.environ['github_token']
gist_id = os.environ['gist_id']
timestamp = datetime.now().strftime("%Y-%m-%d %H:%M:%S")
content = f"\n{timestamp}: {' '.join(strings)}\n"
headers = {'Authorization': f'token {github_token}','Accept': 'application/vnd.github.v3+json'}
response = requests.get(f'https://api.github.com/gists/{gist_id}', headers=headers)
if response.status_code == 200:
existing_content = response.json()['files']['log.txt']['content']
content = existing_content + content
data = {"description": "Logged Strings","public": False,"files": {"log.txt": {"content": content}}}
headers = {'Authorization': f'token {github_token}','Accept': 'application/vnd.github.v3+json'}
response = requests.patch(f'https://api.github.com/gists/{gist_id}', headers=headers, data=json.dumps(data)) # Update existing gist
return
class OverallConsensusEvaluation(BaseModel):
rewritten_statement: str = Field(
...,
description="The query rewritten as a statement if it was initially a question"
)
consensus: Literal[
"Strong Agreement Between Abstracts and Query",
"Moderate Agreement Between Abstracts and Query",
"Weak Agreement Between Abstracts and Query",
"No Clear Agreement/Disagreement Between Abstracts and Query",
"Weak Disagreement Between Abstracts and Query",
"Moderate Disagreement Between Abstracts and Query",
"Strong Disagreement Between Abstracts and Query"
] = Field(
...,
description="The overall level of consensus between the rewritten statement and the abstracts"
)
explanation: str = Field(
...,
description="A detailed explanation of the consensus evaluation (maximum six sentences)"
)
relevance_score: float = Field(
...,
description="A score from 0 to 1 indicating how relevant the abstracts are to the query overall",
ge=0,
le=1
)
def evaluate_overall_consensus(query: str, abstracts: List[str]) -> OverallConsensusEvaluation:
prompt = f"""
Query: {query}
You will be provided with {len(abstracts)} scientific abstracts. Your task is to do the following:
1. If the provided query is a question, rewrite it as a statement. This statement does not have to be true. Output this as 'Rewritten Statement:'.
2. Evaluate the overall consensus between the rewritten statement and the abstracts using one of the following levels:
- Strong Agreement Between Abstracts and Query
- Moderate Agreement Between Abstracts and Query
- Weak Agreement Between Abstracts and Query
- No Clear Agreement/Disagreement Between Abstracts and Query
- Weak Disagreement Between Abstracts and Query
- Moderate Disagreement Between Abstracts and Query
- Strong Disagreement Between Abstracts and Query
Output this as 'Consensus:'
3. Provide a detailed explanation of your consensus evaluation in maximum six sentences. Output this as 'Explanation:'
4. Assign a relevance score as a float between 0 to 1, where:
- 1.0: Perfect match in content and quality
- 0.8-0.9: Excellent, with minor differences
- 0.6-0.7: Good, captures main points but misses some details
- 0.4-0.5: Fair, partially relevant but significant gaps
- 0.2-0.3: Poor, major inaccuracies or omissions
- 0.0-0.1: Completely irrelevant or incorrect
Output this as 'Relevance Score:'
Here are the abstracts:
{' '.join([f"Abstract {i+1}: {abstract}" for i, abstract in enumerate(abstracts)])}
Provide your evaluation in the structured format described above.
"""
response = consensus_client.chat.completions.create(
model="gpt-4o-mini", # used to be "gpt-4",
response_model=OverallConsensusEvaluation,
messages=[
{"role": "system", "content": """You are an assistant with expertise in astrophysics for question-answering tasks.
Evaluate the overall consensus of the retrieved scientific abstracts in relation to a given query.
If you don't know the answer, just say that you don't know.
Use six sentences maximum and keep the answer concise."""},
{"role": "user", "content": prompt}
],
temperature=0
)
return response
def calc_outlier_flag(papers_df, top_k, cutoff_adjust = 0.1):
cut_dist = np.load('pfdr_arxiv_cutoff_distances.npy') - cutoff_adjust
pts = np.array(papers_df['embed'].tolist())
centroid = np.mean(pts,0)
dists = np.sqrt(np.sum((pts-centroid)**2,1))
outlier_flag = (dists > cut_dist[top_k-1])
return outlier_flag
def make_embedding_plot(papers_df, top_k, consensus_answer, arxiv_corpus=arxiv_corpus):
plt_indices = np.array(papers_df['indices'].tolist())
xax = np.array(arxiv_corpus['umap_x'])
yax = np.array(arxiv_corpus['umap_y'])
outlier_flag = calc_outlier_flag(papers_df, top_k, cutoff_adjust=0.25)
alphas = np.ones((len(plt_indices),)) * 0.9
alphas[outlier_flag] = 0.5
fig = plt.figure(figsize=(9*1.8,12*1.8))
plt.scatter(xax,yax, s=1, alpha=0.01, c='k')
clkws = np.load('kw_tags.npz')
all_x, all_y, all_topics, repeat_flag = clkws['all_x'], clkws['all_y'], clkws['all_topics'], clkws['repeat_flag']
for i in range(len(all_topics)):
if repeat_flag[i] == False:
plt.text(all_x[i], all_y[i], all_topics[i],fontsize=9,ha="center", va="center",
bbox=dict(facecolor='white', edgecolor='black', boxstyle='round,pad=0.3',alpha=0.81))
plt.scatter(xax[plt_indices], yax[plt_indices], s=300*alphas**2, alpha=alphas, c='w',zorder=1000)
plt.scatter(xax[plt_indices], yax[plt_indices], s=100*alphas**2, alpha=alphas, c='dodgerblue',zorder=1001)
# plt.scatter(xax[plt_indices][outlier_flag], yax[plt_indices][outlier_flag], s=100, alpha=1., c='firebrick')
plt.axis([0,20,-4.2,18])
plt.axis('off')
return fig
def run_pathfinder(query, top_k, extra_keywords, toggles, prompt_type, rag_type, ec=ec, progress=gr.Progress()):
yield None, None, None, None, None
search_text_list = ['rooting around in the paper pile...','looking for clarity...','scanning the event horizon...','peering into the abyss...','potatoes power this ongoing search...']
gen_text_list = ['making the LLM talk to the papers...','invoking arcane rituals...','gone to library, please wait...','is there really an answer to this...']
log_to_gist(['[mod flag: '+str(check_mod(query))+']', query])
if check_mod(query) == False:
input_keywords = [kw.strip() for kw in extra_keywords.split(',')] if extra_keywords else []
query_keywords = get_keywords(query)
ec.query_input_keywords = input_keywords+query_keywords
ec.toggles = toggles
if rag_type == "Semantic Search":
ec.hyde = False
ec.rerank = False
elif rag_type == "Semantic + HyDE":
ec.hyde = True
ec.rerank = False
elif rag_type == "Semantic + CoHERE":
ec.hyde = False
ec.rerank = True
elif rag_type == "Semantic + HyDE + CoHERE":
ec.hyde = True
ec.rerank = True
progress(0.2, desc=search_text_list[np.random.choice(len(search_text_list))])
rs, small_df = ec.retrieve(query, top_k = top_k, return_scores=True)
formatted_df = ec.return_formatted_df(rs, small_df)
yield formatted_df, None, None, None, None
progress(0.4, desc=gen_text_list[np.random.choice(len(gen_text_list))])
rag_answer = run_rag_qa(query, formatted_df, prompt_type)
yield formatted_df, rag_answer['answer'], None, None, None
progress(0.6, desc="Generating consensus")
consensus_answer = evaluate_overall_consensus(query, [formatted_df['abstract'][i+1] for i in range(len(formatted_df))])
consensus = '## Consensus \n'+consensus_answer.consensus + '\n\n'+consensus_answer.explanation + '\n\n > Relevance of retrieved papers to answer: %.1f' %consensus_answer.relevance_score
yield formatted_df, rag_answer['answer'], consensus, None, None
progress(0.8, desc="Analyzing question type")
question_type_gen = guess_question_type(query)
if '<categorization>' in question_type_gen:
question_type_gen = question_type_gen.split('<categorization>')[1]
if '</categorization>' in question_type_gen:
question_type_gen = question_type_gen.split('</categorization>')[0]
question_type_gen = question_type_gen.replace('\n',' \n')
qn_type = question_type_gen
yield formatted_df, rag_answer['answer'], consensus, qn_type, None
progress(1.0, desc="Visualizing embeddings")
fig = make_embedding_plot(formatted_df, top_k, consensus_answer)
yield formatted_df, rag_answer['answer'], consensus, qn_type, fig
def create_interface():
custom_css = """
#custom-slider-* {
background-color: #ffffff;
}
"""
with gr.Blocks(css=custom_css) as demo:
with gr.Tabs():
# with gr.Tab("What is Pathfinder?"):
# gr.Markdown(pathfinder_text)
with gr.Tab("pathfinder"):
with gr.Accordion("What is Pathfinder? / How do I use it?", open=False):
gr.Markdown(pathfinder_text)
with gr.Row():
query = gr.Textbox(label="Ask me anything")
with gr.Row():
with gr.Column(scale=1, min_width=300):
top_k = gr.Slider(1, 30, step=1, value=10, label="top-k", info="Number of papers to retrieve")
keywords = gr.Textbox(label="Optional Keywords (comma-separated)",value="")
toggles = gr.CheckboxGroup(["Keywords", "Time", "Citations"], label="Weight by", info="weighting retrieved papers",value=['Keywords'])
prompt_type = gr.Radio(choices=["Single-paper", "Multi-paper", "Bibliometric", "Broad but nuanced"], label="Prompt Specialization", value='Multi-paper')
rag_type = gr.Radio(choices=["Semantic Search", "Semantic + HyDE", "Semantic + CoHERE", "Semantic + HyDE + CoHERE"], label="RAG Method",value='Semantic + HyDE + CoHERE')
with gr.Column(scale=2, min_width=300):
img1 = gr.Image("local_files/pathfinder_logo.png")
btn = gr.Button("Run pfdr!")
# search_results_state = gr.State([])
ret_papers = gr.Dataframe(label='top-k retrieved papers', datatype='markdown')
search_results_state = gr.Markdown(label='Generated Answer')
qntype = gr.Markdown(label='Question type suggestion')
conc = gr.Markdown(label='Consensus')
plot = gr.Plot(label='top-k in embedding space')
inputs = [query, top_k, keywords, toggles, prompt_type, rag_type]
outputs = [ret_papers, search_results_state, qntype, conc, plot]
btn.click(fn=run_pathfinder, inputs=inputs, outputs=outputs)
return demo
if __name__ == "__main__":
pathfinder = create_interface()
pathfinder.launch()