synthesist / app.py
kiyer's picture
trying to make things a bit faster
e2d52cc verified
import streamlit as st
st.set_page_config(layout="wide")
import numpy as np
from abc import ABC, abstractmethod
from typing import List, Dict, Any, Tuple
from collections import defaultdict
from tqdm import tqdm
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 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
import instructor
from pydantic import BaseModel, Field
from typing import List, Literal
from nltk.corpus import stopwords
import nltk
from openai import OpenAI
# import anthropic
import cohere
import faiss
import spacy
from string import punctuation
import pytextrank
nlp = spacy.load("en_core_web_sm")
nlp.add_pipe("textrank")
try:
stopwords.words('english')
except:
nltk.download('stopwords')
stopwords.words('english')
from bokeh.plotting import figure
from bokeh.models import ColumnDataSource
from bokeh.io import output_notebook
from bokeh.palettes import Spectral5
from bokeh.transform import linear_cmap
ts = time.time()
# anthropic_key = st.secrets["anthropic_key"]
openai_key = st.secrets["openai_key"]
cohere_key = st.secrets['cohere_key']
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)
st.image('local_files/pathfinder_logo.png')
st.expander("About", expanded=False).write(
"""
Pathfinder v2.0 is a framework for searching and visualizing astronomy papers on the [arXiv](https://arxiv.org/) and [ADS](https://ui.adsabs.harvard.edu/) using the context
sensitivity from modern large language models (LLMs) to better parse patterns in paper contexts.
This tool was built during the [JSALT workshop](https://www.clsp.jhu.edu/2024-jelinek-summer-workshop-on-speech-and-language-technology/) to do awesome things.
**πŸ‘ˆ Use the sidebar to tweak the search parameters to get better results**.
### Tool summary:
- Please wait while the initial data loads and compiles, this takes about a minute initially.
This is not meant to be a replacement to existing tools like the
[ADS](https://ui.adsabs.harvard.edu/),
[arxivsorter](https://www.arxivsorter.org/), semantic search or google scholar, but rather a supplement to find papers
that otherwise might be missed during a literature survey.
It is trained on astro-ph (astrophysics of galaxies) papers up to last-year-ish mined from arxiv and supplemented with ADS metadata,
if you are interested in extending it please reach out!
Also add: feedback form, socials, literature, contact us, copyright, collaboration, etc.
The image below shows a representation of all the astro-ph.GA papers that can be explored in more detail
using the `Arxiv embedding` page. The papers tend to cluster together by similarity, and result in an
atlas that shows well studied (forests) and currently uncharted areas (water).
"""
)
# ---------------- get data and set up session state ---------------------------
if 'arxiv_corpus' not in st.session_state:
with st.spinner('loading data...'):
try:
arxiv_corpus = load_from_disk('data/')
except:
st.write('downloading data')
arxiv_corpus = load_dataset('kiyer/pathfinder_arxiv_data',split='train')
# arxiv_corpus = load_dataset('kiyer/pathfinder_arxiv_data_galaxy',split='train')
arxiv_corpus.save_to_disk('data/')
arxiv_corpus.add_faiss_index('embed')
st.session_state.arxiv_corpus = arxiv_corpus
st.toast('loaded arxiv corpus')
else:
arxiv_corpus = st.session_state.arxiv_corpus
if 'ids' not in st.session_state:
st.session_state.ids = arxiv_corpus['ads_id']
st.session_state.titles = arxiv_corpus['title']
st.session_state.abstracts = arxiv_corpus['abstract']
st.session_state.cites = arxiv_corpus['cites']
st.session_state.years = arxiv_corpus['date']
st.session_state.kws = arxiv_corpus['keywords']
st.session_state.ads_kws = arxiv_corpus['ads_keywords']
st.session_state.bibcode = arxiv_corpus['bibcode']
st.session_state.umap_x = arxiv_corpus['umap_x']
st.session_state.umap_y = arxiv_corpus['umap_y']
st.toast('done caching. time taken: %.2f sec' %(time.time()-ts))
#---------------------------------------------------------------
# A hack to "clear" the previous result when submitting a new prompt. This avoids
# the "previous run's text is grayed-out but visible during rerun" Streamlit behavior.
class DirtyState:
NOT_DIRTY = "NOT_DIRTY"
DIRTY = "DIRTY"
UNHANDLED_SUBMIT = "UNHANDLED_SUBMIT"
def get_dirty_state() -> str:
return st.session_state.get("dirty_state", DirtyState.NOT_DIRTY)
def set_dirty_state(state: str) -> None:
st.session_state["dirty_state"] = state
def with_clear_container(submit_clicked: bool) -> bool:
if get_dirty_state() == DirtyState.DIRTY:
if submit_clicked:
set_dirty_state(DirtyState.UNHANDLED_SUBMIT)
st.experimental_rerun()
else:
set_dirty_state(DirtyState.NOT_DIRTY)
if submit_clicked or get_dirty_state() == DirtyState.UNHANDLED_SUBMIT:
set_dirty_state(DirtyState.DIRTY)
return True
return False
# ---------------- define embedding retrieval systems --------------------------
def get_keywords(text):
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 parse_doc(text, nret = 10):
local_kws = []
doc = nlp(text)
# examine the top-ranked phrases in the document
for phrase in doc._.phrases[:nret]:
# print(phrase.text)
local_kws.append(phrase.text)
return local_kws
class EmbeddingRetrievalSystem():
def __init__(self, weight_citation = False, weight_date = False, weight_keywords = False):
self.ids = st.session_state.ids
self.years = st.session_state.years
self.abstract = st.session_state.abstracts
self.client = OpenAI(api_key = openai_key)
self.embed_model = "text-embedding-3-small"
self.dataset = st.session_state.arxiv_corpus
self.kws = st.session_state.kws
self.cites = st.session_state.cites
self.weight_citation = weight_citation
self.weight_date = weight_date
self.weight_keywords = weight_keywords
self.id_to_index = {self.ids[i]: i for i in range(len(self.ids))}
# self.citation_filter = CitationFilter(self.dataset)
# self.date_filter = DateFilter(self.dataset['date'])
# self.keyword_filter = KeywordFilter(corpus=self.dataset, remove_capitals=True)
def parse_date(self, id):
# indexval = np.where(self.ids == id)[0][0]
indexval = id
return self.years[indexval]
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 analyze_temporal_query(self, query):
return
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]
def rank_and_filter(self, query, query_embedding, query_date, top_k = 10, return_scores=False, time_result=None):
# st.write('status')
# st.write('toggles', self.toggles)
# st.write('question_type', self.question_type)
# st.write('rag method', self.rag_method)
# st.write('gen method', self.gen_method)
self.weight_keywords = self.toggles["Keyword weighting"]
self.weight_date = self.toggles["Time weighting"]
self.weight_citation = self.toggles["Citation weighting"]
topk_indices, similarities = self.calc_faiss(np.array(query_embedding), top_k = 1000)
similarities = 1/similarities # converting from a distance (less is better) to a similarity (more is better)
query_kws = get_keywords(query)
input_kws = self.query_input_keywords
query_kws = query_kws + input_kws
self.query_kws = query_kws
if self.weight_keywords == True:
sub_kws = [self.kws[i] for i in topk_indices]
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 = [self.years[i] for i in topk_indices]
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([self.cites[i] for i in topk_indices])
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)
# if self.weight_keywords:
# keyword_matches = self.keyword_filter.filter(query)
# self.query_kws = keyword_matches
# kw_indices = np.zeros_like(similarities)
# for s in keyword_matches:
# if self.id_to_index[s] in topk_indices:
# # print('yes', self.id_to_index[s], topk_indices[np.where(topk_indices == self.id_to_index[s])[0]])
# similarities[np.where(topk_indices == self.id_to_index[s])[0]] = similarities[np.where(topk_indices == self.id_to_index[s])[0]] * 10.
# similarities = similarities / 10.
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]
if return_scores:
return {doc[0]: doc[1] for doc in top_results}
# Only keep the document IDs
top_results = [doc[0] for doc in top_results]
return top_results
def retrieve(self, query, top_k, time_result=None, query_date = None, return_scores = False):
query_embedding = self.get_query_embedding(query)
# Judge time relevance
if time_result is None:
if self.weight_date:
time_result, time_taken = self.analyze_temporal_query(query, self.anthropic_client)
else:
time_result = {'has_temporal_aspect': False, 'expected_year_filter': None, 'expected_recency_weight': None}
top_results = self.rank_and_filter(query,
query_embedding,
query_date,
top_k,
return_scores = return_scores,
time_result = time_result)
return top_results
class HydeRetrievalSystem(EmbeddingRetrievalSystem):
def __init__(self, generation_model: str = "claude-3-haiku-20240307",
embedding_model: str = "text-embedding-3-small",
temperature: float = 0.5,
max_doclen: int = 500,
generate_n: int = 1,
embed_query = True,
conclusion = False, **kwargs):
# Handle the kwargs for the superclass init -- filters/citation weighting
super().__init__(**kwargs)
if max_doclen * generate_n > 8191:
raise ValueError("Too many tokens. Please reduce max_doclen or generate_n.")
self.embedding_model = embedding_model
self.generation_model = generation_model
# HYPERPARAMETERS
self.temperature = temperature # generation temperature
self.max_doclen = max_doclen # max tokens for generation
self.generate_n = generate_n # how many documents
self.embed_query = embed_query # embed the query vector?
self.conclusion = conclusion # generate conclusion as well?
# self.anthropic_key = anthropic_key
# self.generation_client = anthropic.Anthropic(api_key = self.anthropic_key)
self.generation_client = openai_llm(temperature=0,model_name='gpt-4o-mini', openai_api_key = openai_key)
def retrieve(self, query: str, top_k: int = 10, return_scores = False, time_result = None) -> List[Tuple[str, str, float]]:
if time_result is None:
if self.weight_date: time_result, time_taken = analyze_temporal_query(query, self.anthropic_client)
else: time_result = {'has_temporal_aspect': False, 'expected_year_filter': None, 'expected_recency_weight': None}
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)
embedding = np.mean(np.array(doc_embeddings), axis = 0)
top_results = self.rank_and_filter(query, embedding, query_date=None, top_k = top_k, return_scores = return_scores, time_result = time_result)
return top_results
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)
# st.write('invoking hyde generation')
# message = self.generation_client.messages.create(
# model = self.generation_model,
# max_tokens = self.max_doclen,
# temperature = self.temperature,
# system = prompt,
# messages=[{ "role": "user",
# "content": [{"type": "text", "text": query,}] }]
# )
# return message.content[0].text
messages = [("system",prompt,),("human", query),]
return self.generation_client.invoke(messages).content
def generate_docs(self, query: str):
docs = []
for i in range(self.generate_n):
# st.write('invoking hyde generation2')
docs.append(self.generate_doc(query))
# with concurrent.futures.ThreadPoolExecutor() as executor:
# st.write('invoking hyde generation2')
# future_to_query = {executor.submit(self.generate_doc, query): query for i in range(self.generate_n)}
# for future in concurrent.futures.as_completed(future_to_query):
# query = future_to_query[future]
# try:
# data = future.result()
# docs.append(data)
# except Exception as exc:
# pass
return docs
def embed_docs(self, docs: List[str]):
return self.embed_batch(docs)
class HydeCohereRetrievalSystem(HydeRetrievalSystem):
def __init__(self, **kwargs):
super().__init__(**kwargs)
self.cohere_key = cohere_key
self.cohere_client = cohere.Client(self.cohere_key)
def retrieve(self, query: str,
top_k: int = 10,
rerank_top_k: int = 250,
return_scores = False, time_result = None,
reweight = False) -> List[Tuple[str, str, float]]:
if time_result is None:
if self.weight_date: time_result, time_taken = analyze_temporal_query(query, self.anthropic_client)
else: time_result = {'has_temporal_aspect': False, 'expected_year_filter': None, 'expected_recency_weight': None}
top_results = super().retrieve(query, top_k = rerank_top_k, time_result = time_result)
# doc_texts = self.get_document_texts(top_results)
# docs_for_rerank = [f"Abstract: {doc['abstract']}\nConclusions: {doc['conclusions']}" for doc in doc_texts]
docs_for_rerank = [self.abstract[i] for i in top_results]
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])
if reweight:
if time_result['has_temporal_aspect']:
final_results = self.date_filter.filter(final_results, time_score = time_result['expected_recency_weight'])
if self.weight_citation: self.citation_filter.filter(final_results)
if return_scores:
return {result[0]: result[2] for result in final_results}
return [doc[0] for doc in final_results]
def embed_docs(self, docs: List[str]):
return self.embed_batch(docs)
# ----------------------------------------------------------------
if 'ec' not in st.session_state:
ec = HydeCohereRetrievalSystem(weight_keywords=True)
st.session_state.ec = ec
st.toast('loaded retrieval system')
else:
ec = st.session_state.ec
def get_topk(query, top_k):
print('running retrieval')
rs = st.session_state.ec.retrieve(query, top_k, return_scores=True)
return rs
def Library(query, top_k = 7):
rs = get_topk(query, top_k = top_k)
op_docs = ''
for paperno, i in enumerate(rs):
op_docs = op_docs + 'Paper %.0f:' %(paperno+1) +' (published in '+st.session_state.bibcode[i][0:4] + ') ' + st.session_state.titles[i] + '\n' + st.session_state.abstracts[i] + '\n\n'
return op_docs
def Library2(query, top_k = 7):
rs = get_topk(query, top_k = top_k)
absts, fnames = [], []
for paperno, i in enumerate(rs):
absts.append(st.session_state.abstracts[i])
fnames.append(st.session_state.bibcode[i])
return absts, fnames, rs
def get_paper_df(ids):
papers, scores, yrs, links, cites, kws = [], [], [], [], [], []
for i in ids:
papers.append(st.session_state.titles[i])
scores.append(ids[i])
links.append('https://ui.adsabs.harvard.edu/abs/'+st.session_state.bibcode[i]+'/abstract')
yrs.append(st.session_state.bibcode[i][0:4])
cites.append(st.session_state.cites[i])
kws.append(st.session_state.ads_kws[i])
return pd.DataFrame({
'Title': papers,
'Relevance': scores,
'Year': yrs,
'ADS Link': links,
'Citations': cites,
'Keywords': kws,
})
# def find_outliers(inp_simids, arxiv_cutoff_distance = 0.8):
#
# inp_simids = np.array(inp_simids)
#
# # Calculate the centroid for each point, excluding itself
# orange_black_points = st.session_state.embed[inp_simids]
#
# topk_dists = []
# for i, point in enumerate(orange_black_points):
# # Exclude the current point
# other_points = np.delete(orange_black_points, i, axis=0)
# # Calculate centroid of other points
# centroid = np.mean(other_points, axis=0)
# # Calculate distance from the point to this centroid
# dist = np.sqrt(np.sum((point - centroid)**2))
# topk_dists.append(dist)
#
# topk_dists = np.array(topk_dists)
#
# # Separate distances for orange and black points
# orange_distances = topk_dists[:len(inp_simids)]
# black_distances = topk_dists[len(inp_simids):]
#
# # Calculate the median of distances
# orange_black_distances = topk_dists
# median_topk_distance = np.median(orange_black_distances)
#
# # def get_sims_and_dists(inp_data):
#
# # all_sims, all_dists = [], []
#
# # np.random.seed(12)
# # rand_indices = np.random.choice(inp_data.shape[0], size=return_n, replace=False)
#
# # for j in tqdm(range(len(rand_indices))):
#
# # i = rand_indices[j]
# # inferred_vector = inp_data[i,:]
# # sims, dists = find_closest_dists(i, inp_data, return_n + 1)
# # all_sims.append(sims[1:])
# # all_dists.append(dists[1:])
#
# # return np.array(all_sims), np.array(all_dists)
#
# # # Identify papers with distances greater than the 95th percentile
# # _, all_dists = get_sims_and_dists(arxiv_ada_embeddings)
# # arxiv_cutoff_distance = find_cutoff_dist(all_dists)
# # hardcoding for now
# outlier_indices = inp_simids[np.where(orange_black_distances > arxiv_cutoff_distance)[0]]
# # outlier_titles = [titles[i] for i in outlier_indices]
#
# return outlier_indices #, outlier_titles
def create_embedding_plot(rs):
"""
function to create embedding plot
"""
pltsource = ColumnDataSource(data=dict(
x=st.session_state.umap_x,
y=st.session_state.umap_y,
title=st.session_state.titles,
link=st.session_state.bibcode,
))
rsflag = np.zeros((len(st.session_state.ids),))
rsflag[np.array([k for k in rs])] = 1
# outflag = np.zeros((len(st.session_state.ids),))
# outflag[np.array([k for k in find_outliers(rs)])] = 1
pltsource.data['colors'] = rsflag * 0.8 + 0.1
# pltsource.data['colors'][outflag] = 0.5
pltsource.data['sizes'] = (rsflag + 1)**5 / 100
TOOLTIPS = """
<div style="width:300px;">
ID: $index
($x, $y)
@title <br>
@link <br> <br>
</div>
"""
mapper = linear_cmap(field_name="colors", palette=Spectral5, low=0., high=1.)
p = figure(width=700, height=900, tooltips=TOOLTIPS, x_range=(0, 20), y_range=(-4.2,18),
title="UMAP projection of embeddings for the astro-ph corpus")
p.axis.visible=False
p.grid.visible=False
p.outline_line_alpha = 0.
p.circle('x', 'y', radius='sizes', source=pltsource, alpha=0.3, fill_color=mapper, fill_alpha='colors', line_color="lightgrey",line_alpha=0.1)
return p
def extract_keywords(question, ec):
# Simulated keyword extraction (replace with actual logic)
return ['keyword1', 'keyword2', 'keyword3']
# Function to estimate consensus (replace with actual implementation)
def estimate_consensus():
# Simulated consensus estimation (replace with actual calculation)
return 0.75
def run_agent_qa(query, top_k):
# define tools
search = DuckDuckGoSearchAPIWrapper()
tools = [
Tool(
name="Library",
func=Library,
description="A source of information pertinent to your question. Do not answer a question without consulting this!"
),
Tool(
name="Search",
func=search.run,
description="useful for when you need to look up knowledge about common topics or current events",
)
]
if 'tools' not in st.session_state:
st.session_state.tools = tools
# define prompt
# for another question type:
# First, find the quotes from the document that are most relevant to answering the question, and then print them in numbered order.
# Quotes should be relatively short. If there are no relevant quotes, write β€œNo relevant quotes” instead.
template = """You are an expert astronomer and cosmologist.
Answer the following question as best you can using information from the library, but speaking in a concise and factual manner.
If you can not come up with an answer, say you do not know.
Try to break the question down into smaller steps and solve it in a logical manner.
You have access to the following tools:
{tools}
Use the following format:
Question: the input question you must answer
Thought: you should always think about what to do
Action: the action to take, should be one of [{tool_names}]
Action Input: the input to the action
Observation: the result of the action
... (this Thought/Action/Action Input/Observation can repeat N times)
Thought: I now know the final answer
Final Answer: the final answer to the original input question. provide information about how you arrived at the answer, and any nuances or uncertainties the reader should be aware of
Begin! Remember to speak in a pedagogical and factual manner."
Question: {input}
Thought:{agent_scratchpad}"""
prompt = hub.pull("hwchase17/react")
prompt.template=template
# path to write intermediate trace to
file_path = "agent_trace.txt"
try:
os.remove(file_path)
except:
pass
file_handler = FileCallbackHandler(file_path)
callback_manager=CallbackManager([file_handler])
# define and execute agent
tool_names = [tool.name for tool in st.session_state.tools]
if 'agent' not in st.session_state:
# agent = ZeroShotAgent(llm_chain=llm_chain, allowed_tools=tool_names)
agent = create_react_agent(llm=gen_llm, tools=tools, prompt=prompt)
st.session_state.agent = agent
if 'agent_executor' not in st.session_state:
agent_executor = AgentExecutor(agent=st.session_state.agent, tools=st.session_state.tools, verbose=True, handle_parsing_errors=True, callbacks=CallbackManager([file_handler]))
st.session_state.agent_executor = agent_executor
answer = st.session_state.agent_executor.invoke({"input": query,})
return answer
def make_rag_qa_answer(query, top_k = 10):
absts, fhdrs, rs = Library2(query, top_k = top_k)
temp_abst = ''
loaders = []
for i in range(len(absts)):
temp_abst = absts[i]
try:
text_file = open("absts/"+fhdrs[i]+".txt", "w")
except:
os.mkdir('absts')
text_file = open("absts/"+fhdrs[i]+".txt", "w")
n = text_file.write(temp_abst)
text_file.close()
loader = TextLoader("absts/"+fhdrs[i]+".txt")
loaders.append(loader)
text_splitter = RecursiveCharacterTextSplitter(chunk_size=150, chunk_overlap=50, add_start_index=True)
splits = text_splitter.split_documents([loader.load()[0] for loader in loaders])
vectorstore = Chroma.from_documents(documents=splits, embedding=embeddings, collection_name='retdoc4')
# retriever = vectorstore.as_retriever(search_type="similarity", search_kwargs={"k": 6, "fetch_k": len(splits)})
retriever = vectorstore.as_retriever(search_type="similarity", search_kwargs={"k": 6})
for i in range(len(absts)):
os.remove("absts/"+fhdrs[i]+".txt")
template = """You are an expert astronomer and cosmologist.
Answer the following question as best you can using information from the library, but speaking in a concise and factual manner.
If you can not come up with an answer, say you do not know.
Try to break the question down into smaller steps and solve it in a logical manner.
Provide information about how you arrived at the answer, and any nuances or uncertainties the reader should be aware of.
Begin! Remember to speak in a pedagogical and factual manner."
Relevant documents:{context}
Question: {question}
Answer:"""
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()
return rag_answer, rs
def guess_question_type(query: str):
categorization_prompt = """You are an expert astrophysicist and computer scientist specializing in linguistics and semantics. Your task is to categorize a given query into one of the following categories:
1. Summarization
2. Single-paper factual
3. Multi-paper factual
4. Named entity recognition
5. Jargon-specific questions / overloaded words
6. Time-sensitive
7. Consensus evaluation
8. What-ifs and counterfactuals
9. Compositional
Analyze the query carefully, considering its content, structure, and implications. Then, determine which of the above categories best fits the query.
In your analysis, consider the following:
- Does the query ask for a well-known datapoint or mechanism?
- Can it be answered by a single paper or does it require multiple sources?
- Does it involve proper nouns or specific scientific terms?
- Is it time-dependent or likely to change in the near future?
- Does it require evaluating consensus across multiple sources?
- Is it a hypothetical or counterfactual question?
- Does it need to be broken down into sub-queries (i.e. compositional)?
After your analysis, categorize the query into one of the nine categories listed above.
Provide a brief explanation for your categorization, highlighting the key aspects of the query that led to your decision.
Present your final answer in the following format:
<categorization>
Category: [Selected category]
Explanation: [Your explanation for the categorization]
</categorization>"""
# st.write('invoking hyde generation')
# message = self.generation_client.messages.create(
# model = self.generation_model,
# max_tokens = self.max_doclen,
# temperature = self.temperature,
# system = prompt,
# messages=[{ "role": "user",
# "content": [{"type": "text", "text": query,}] }]
# )
# return message.content[0].text
messages = [("system",categorization_prompt,),("human", query),]
return st.session_state.ec.generation_client.invoke(messages).content
class OverallConsensusEvaluation(BaseModel):
consensus: Literal["Strong Agreement", "Moderate Agreement", "Weak Agreement", "No Clear Consensus", "Weak Disagreement", "Moderate Disagreement", "Strong Disagreement"] = Field(
...,
description="The overall level of consensus between the query and the abstracts"
)
explanation: str = Field(
...,
description="A detailed explanation of the consensus evaluation"
)
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:
"""
Evaluates the overall consensus of the abstracts in relation to the query in a single LLM call.
"""
prompt = f"""
Query: {query}
You will be provided with {len(abstracts)} scientific abstracts. Your task is to:
1. Evaluate the overall consensus between the query and the abstracts.
2. Provide a detailed explanation of your consensus evaluation.
3. Assign an overall relevance score from 0 to 1, where 0 means completely irrelevant and 1 means highly relevant.
For the consensus evaluation, use one of the following levels:
Strong Agreement, Moderate Agreement, Weak Agreement, No Clear Consensus, Weak Disagreement, Moderate Disagreement, Strong Disagreement
Here are the abstracts:
{' '.join([f"Abstract {i+1}: {abstract}" for i, abstract in enumerate(abstracts)])}
Provide your evaluation in a structured format.
"""
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
# Streamlit app
def main():
# st.title("Question Answering App")
# Sidebar (Inputs)
st.sidebar.header("Fine-tune the search")
top_k = st.sidebar.slider("Number of papers to retrieve:", 3, 30, 10)
extra_keywords = st.sidebar.text_input("Enter extra keywords (comma-separated):")
st.sidebar.subheader("Toggles")
toggle_a = st.sidebar.toggle("Weight by keywords", value = False)
toggle_b = st.sidebar.toggle("Weight by date", value = False)
toggle_c = st.sidebar.toggle("Weight by citations", value = False)
method = st.sidebar.radio("Retrieval method:", ["Semantic search", "Semantic search + HyDE", "Semantic search + HyDE + CoHERE"], index=2)
if (method == "Semantic search"):
with st.spinner('set retrieval method to'+ method):
st.session_state.ec = EmbeddingRetrievalSystem(weight_keywords=True)
elif (method == "Semantic search + HyDE"):
with st.spinner('set retrieval method to'+ method):
st.session_state.ec = HydeRetrievalSystem(weight_keywords=True)
elif (method == "Semantic search + HyDE + CoHERE"):
with st.spinner('set retrieval method to'+ method):
st.session_state.ec = HydeCohereRetrievalSystem(weight_keywords=True)
method2 = st.sidebar.radio("Generation complexity:", ["Basic RAG","ReAct Agent"])
if method2 == "Basic RAG":
st.session_state.gen_method = 'rag'
elif method2 == "ReAct Agent":
st.session_state.gen_method = 'agent'
question_type = st.sidebar.selectbox("Select question type:", ["Single paper", "Multi-paper", "Summary"])
store_output = st.sidebar.button("Save output")
# Main page (Outputs)
# st.markdown("""
# <style>
# .stTextInput > div > div { font-size: 50px; }
# </style>
# """, unsafe_allow_html=True)
# st.markdown(
# """
# <style>
# textarea {
# font-size: 3rem !important;
# font-weight: bold;
# font-family: "Times New Roman", Times, serif;
# }
# input {
# font-size: 3rem !important;
# font-weight: bold;
# font-family: "Times New Roman", Times, serif;
# }
# </style>
# """,
# unsafe_allow_html=True,
# )
# query = st.text_area("Ask me anything:", height=30)
query = st.text_input("Ask me anything:")
submit_button = st.button("Submit")
if submit_button:
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...']
with st.spinner(search_text_list[np.random.choice(len(search_text_list))]):
# Process inputs
keywords = [kw.strip() for kw in extra_keywords.split(',')] if extra_keywords else []
toggles = {'Keyword weighting': toggle_a, 'Time weighting': toggle_b, 'Citation weighting': toggle_c}
# Generate outputs
st.session_state.ec.query_input_keywords = keywords
st.session_state.ec.toggles = toggles
st.session_state.ec.question_type = question_type
st.session_state.ec.rag_method = method
st.session_state.ec.gen_method = method2
# Display outputs
if st.session_state.gen_method == 'agent':
answer = run_agent_qa(query, top_k)
rs = get_topk(query, top_k)
st.write(answer["output"])
file_path = "agent_trace.txt"
with open(file_path, 'r') as file:
intermediate_steps = file.read()
st.expander('Intermediate steps', expanded=False).write(intermediate_steps)
elif st.session_state.gen_method == 'rag':
answer, rs = make_rag_qa_answer(query, top_k)
st.write(answer['answer'])
papers_df = get_paper_df(rs)
embedding_plot = create_embedding_plot(rs)
triggered_keywords = st.session_state.ec.query_kws
st.write('**Triggered keywords:** `'+ "`, `".join(triggered_keywords)+'`')
# consensus = estimate_consensus()
with st.expander("Relevant papers", expanded=True):
# st.dataframe(papers_df, hide_index=True)
st.data_editor(papers_df,
column_config = {'ADS Link':st.column_config.LinkColumn(display_text= 'https://ui.adsabs.harvard.edu/abs/(.*?)/abstract')}
)
# with st.expander("Embedding map", expanded=False):
st.bokeh_chart(embedding_plot)
col1, col2 = st.columns(2)
with col1:
st.subheader("Question type suggestion")
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')
st.markdown(question_type_gen)
with col2:
# st.subheader("Triggered Keywords")
# st.write(", ".join(triggered_keywords))
consensus_answer = evaluate_overall_consensus(query, [st.session_state.abstracts[i] for i in rs])
st.subheader("Consensus: "+consensus_answer.consensus)
st.markdown(consensus_answer.explanation)
st.markdown('Relevance of retrieved papers to answer: %.1f' %consensus_answer.relevance_score)
# st.write(f"{consensus:.2%}")
else:
st.info("Use the sidebar to tweak the search parameters to get better results.")
if store_output:
st.toast("Output stored successfully!")
if __name__ == "__main__":
main()