synthesist / fns.py
kiyer's picture
basic files and codebase
6931cbb verified
import time
s2 = time.time()
import numpy as np
import streamlit as st
import json
from abc import ABC, abstractmethod
from typing import List, Dict, Any, Tuple
from collections import defaultdict
# import wandb
import numpy as np
from tqdm import tqdm
from datetime import datetime, date
import pickle
from datasets import load_dataset
import os
from nltk.corpus import stopwords
import nltk
from openai import OpenAI
import anthropic
import time
from collections import Counter
try:
stopwords.words('english')
except:
nltk.download('stopwords')
stopwords.words('english')
openai_key = st.secrets['openai_key']
anthropic_key = st.secrets['anthropic_key']
# anthropic_key = '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'
@st.cache_data
def load_astro_meta():
print('load astro meta')
return load_dataset('arxiv_corpus/', split = "train")
@st.cache_data
def load_index_mapping(index_mapping_path):
print("Loading index mapping...")
with open(index_mapping_path, 'rb') as f:
temp = pickle.load(f)
return temp
@st.cache_data
def load_embeddings(embeddings_path):
print("Loading embedding")
return np.load(embeddings_path)
@st.cache_data
def load_metadata(meta_path):
print("Loading metadata...")
with open(meta_path, 'r') as f:
metadata = json.load(f)
return metadata
# @st.cache_data
def load_umapcoords(umap_path):
print('loading umap coords')
with open(umap_path, "rb") as fp: #Pickling
umap = pickle.load(fp)
return umap
class EmbeddingClient:
def __init__(self, client: OpenAI, model: str = "text-embedding-3-small"):
self.client = client
self.model = model
def embed(self, text: str) -> np.ndarray:
embedding = self.client.embeddings.create(input=[text], model=self.model).data[0].embedding
return np.array(embedding, dtype=np.float32)
def embed_batch(self, texts: List[str]) -> List[np.ndarray]:
embeddings = self.client.embeddings.create(input=texts, model=self.model).data
return [np.array(embedding.embedding, dtype=np.float32) for embedding in embeddings]
class RetrievalSystem(ABC):
@abstractmethod
def retrieve(self, query: str, arxiv_id: str, top_k: int = 100) -> List[str]:
pass
def parse_date(self, arxiv_id: str) -> datetime:
if arxiv_id is None:
return date.today()
if arxiv_id.startswith('astro-ph'):
arxiv_id = arxiv_id.split('astro-ph')[1].split('_arXiv')[0]
try:
year = int("20" + arxiv_id[:2])
month = int(arxiv_id[2:4])
except:
year = 2023
month = 1
return date(year, month, 1)
class EmbeddingRetrievalSystem(RetrievalSystem):
def __init__(self, embeddings_path: str = "local_files/embeddings_matrix.npy",
documents_path: str = "local_files/documents.pkl",
index_mapping_path: str = "local_files/index_mapping.pkl",
metadata_path: str = "local_files/metadata.json",
weight_citation = False, weight_date = False, weight_keywords = False):
self.embeddings_path = embeddings_path
self.documents_path = documents_path
self.index_mapping_path = index_mapping_path
self.metadata_path = metadata_path
self.weight_citation = weight_citation
self.weight_date = weight_date
self.weight_keywords = weight_keywords
self.embeddings = None
self.documents = None
self.index_mapping = None
self.metadata = None
self.document_dates = []
self.load_data()
self.init_filters()
# config = yaml.safe_load(open('../config.yaml', 'r'))
self.client = EmbeddingClient(OpenAI(api_key=openai_key))
self.anthropic_client = anthropic.Anthropic(api_key=anthropic_key)
def generate_metadata(self):
astro_meta = load_astro_meta()
# dataset = load_dataset('arxiv_corpus/')
keys = list(astro_meta[0].keys())
keys.remove('abstract')
keys.remove('introduction')
keys.remove('conclusions')
self.metadata = {}
for paper in astro_meta:
id_str = paper['arxiv_id']
self.metadata[id_str] = {key: paper[key] for key in keys}
with open(self.metadata_path, 'w') as f:
json.dump(self.metadata, f)
st.markdown("Wrote metadata to {}".format(self.metadata_path))
#
def load_data(self):
# print("Loading embeddings...")
# self.embeddings = np.load(self.embeddings_path)
self.embeddings = load_embeddings(self.embeddings_path)
st.sidebar.success("Loaded embeddings")
# with open(self.index_mapping_path, 'rb') as f:
# self.index_mapping = pickle.load(f)
self.index_mapping = load_index_mapping(self.index_mapping_path)
st.sidebar.success("Loaded index mapping")
# print("Loading documents...")
# with open(self.documents_path, 'rb') as f:
# self.documents = pickle.load(f)
dataset = load_astro_meta()
st.sidebar.success("Loaded documents")
print("Processing document dates...")
# self.document_dates = {doc.id: self.parse_date(doc.arxiv_id) for doc in self.documents}
aids = dataset['arxiv_id']
adsids = dataset['id']
self.document_dates = {adsids[i]: self.parse_date(aids[i]) for i in range(len(aids))}
if os.path.exists(self.metadata_path):
self.metadata = load_metadata(self.metadata_path)
print("Loaded metadata.")
else:
print("Could not find path; generating metadata.")
self.generate_metadata()
print("Data loaded successfully.")
def init_filters(self):
print("Loading filters...")
self.citation_filter = CitationFilter(metadata = self.metadata)
self.date_filter = DateFilter(document_dates = self.document_dates)
self.keyword_filter = KeywordFilter(index_path = "local_files/keyword_index.json", metadata = self.metadata, remove_capitals = True)
def retrieve(self, query: str, arxiv_id: str = None, top_k: int = 10, return_scores = False, time_result = None) -> List[Tuple[str, str, float]]:
query_date = self.parse_date(arxiv_id)
query_embedding = self.get_query_embedding(query)
# Judge time relevance
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 = self.rank_and_filter(query, query_embedding, query_date, top_k, return_scores = return_scores, time_result = time_result)
return top_results
def rank_and_filter(self, query, query_embedding: np.ndarray, query_date, top_k: int = 10, return_scores = False, time_result = None) -> List[Tuple[str, str, float]]:
# Calculate similarities
similarities = np.dot(self.embeddings, query_embedding)
# Filter and rank results
if self.weight_keywords: keyword_matches = self.keyword_filter.filter(query)
results = []
for doc_id, mappings in self.index_mapping.items():
if not self.weight_keywords or doc_id in keyword_matches:
abstract_sim = similarities[mappings['abstract']] if 'abstract' in mappings else -np.inf
conclusions_sim = similarities[mappings['conclusions']] if 'conclusions' in mappings else -np.inf
if abstract_sim > conclusions_sim:
results.append([doc_id, "abstract", abstract_sim])
else:
results.append([doc_id, "conclusions", conclusions_sim])
# Sort and weight and get top-k results
if time_result['has_temporal_aspect']:
filtered_results = self.date_filter.filter(results, boolean_date = time_result['expected_year_filter'], time_score = time_result['expected_recency_weight'], max_date = query_date)
else:
filtered_results = self.date_filter.filter(results, max_date = query_date)
if self.weight_citation: self.citation_filter.filter(filtered_results)
top_results = sorted(filtered_results, key=lambda x: x[2], reverse=True)[:top_k]
if return_scores:
return {doc[0]: doc[2] for doc in top_results}
# Only keep the document IDs
top_results = [doc[0] for doc in top_results]
return top_results
def get_query_embedding(self, query: str) -> np.ndarray:
embedding = self.client.embed(query)
return np.array(embedding, dtype = np.float32)
def get_document_texts(self, doc_ids: List[str]) -> List[Dict[str, str]]:
results = []
for doc_id in doc_ids:
doc = next((d for d in self.documents if d.id == doc_id), None)
if doc:
results.append({
'id': doc.id,
'abstract': doc.abstract,
'conclusions': doc.conclusions
})
else:
print(f"Warning: Document with ID {doc_id} not found.")
return results
def retrieve_context(self, query, top_k, sections = ["abstract", "conclusions"], **kwargs):
docs = self.retrieve(query, top_k = top_k, return_scores = True, **kwargs)
docids = docs.keys()
doctexts = self.get_document_texts(docids) # avoid having to do this repetitively?
context_str = ""
doclist = []
for docid, doctext in zip(docids, doctexts):
for section in sections:
context_str += f"{docid}: {doctext[section]}\n"
meta_row = self.metadata[docid]
doclist.append(Document(docid, doctext['abstract'], doctext['conclusions'], docid, title = meta_row['title'],
score = docs[docid], n_citation = meta_row['citation_count'], keywords = meta_row['keyword_search']))
return context_str, doclist
class Filter():
def filter(self, query: str, arxiv_id: str) -> List[str]:
pass
class CitationFilter(Filter): # can do it with all metadata
def __init__(self, metadata):
self.metadata = metadata
self.citation_counts = {doc_id: self.metadata[doc_id]['citation_count'] for doc_id in self.metadata}
def citation_weight(self, x, shift, scale):
return 1 / (1 + np.exp(-1 * (x - shift) / scale)) # sigmoid function
def filter(self, doc_scores, weight = 0.1): # additive weighting
citation_count = np.array([self.citation_counts[doc[0]] for doc in doc_scores])
cmean, cstd = np.median(citation_count), np.std(citation_count)
citation_score = self.citation_weight(citation_count, cmean, cstd)
for i, doc in enumerate(doc_scores):
doc_scores[i][2] += weight * citation_score[i]
class DateFilter(Filter): # include time weighting eventually
def __init__(self, document_dates):
self.document_dates = document_dates
def parse_date(self, arxiv_id: str) -> datetime: # only for documents
if arxiv_id.startswith('astro-ph'):
arxiv_id = arxiv_id.split('astro-ph')[1].split('_arXiv')[0]
try:
year = int("20" + arxiv_id[:2])
month = int(arxiv_id[2:4])
except:
year = 2023
month = 1
return date(year, month, 1)
def weight(self, time, shift, scale):
return 1 / (1 + np.exp((time - shift) / scale))
def evaluate_filter(self, year, filter_string):
try:
# Use ast.literal_eval to safely evaluate the expression
result = eval(filter_string, {"__builtins__": None}, {"year": year})
return result
except Exception as e:
print(f"Error evaluating filter: {e}")
return False
def filter(self, docs, boolean_date = None, min_date = None, max_date = None, time_score = 0):
filtered = []
if boolean_date is not None:
boolean_date = boolean_date.replace("AND", "and").replace("OR", "or")
for doc in docs:
if self.evaluate_filter(self.document_dates[doc[0]].year, boolean_date):
filtered.append(doc)
else:
if min_date == None: min_date = date(1990, 1, 1)
if max_date == None: max_date = date(2024, 7, 3)
for doc in docs:
if self.document_dates[doc[0]] >= min_date and self.document_dates[doc[0]] <= max_date:
filtered.append(doc)
if time_score is not None: # apply time weighting
for i, item in enumerate(filtered):
time_diff = (max_date - self.document_dates[filtered[i][0]]).days / 365
filtered[i][2] += time_score * 0.1 * self.weight(time_diff, 5, 5)
return filtered
class KeywordFilter(Filter):
def __init__(self, index_path: str = "local_files/keyword_index.json",
remove_capitals: bool = True, metadata = None, ne_only = True, verbose = False):
self.index_path = index_path
self.metadata = metadata
self.remove_capitals = remove_capitals
self.ne_only = ne_only
self.stopwords = set(stopwords.words('english'))
self.verbose = verbose
self.index = None
self.load_or_build_index()
def preprocess_text(self, text: str) -> str:
text = ''.join(char for char in text if char.isalnum() or char.isspace())
if self.remove_capitals: text = text.lower()
return ' '.join(word for word in text.split() if word.lower() not in self.stopwords)
def build_index(self): # include the title in the index
print("Building index...")
self.index = {}
for i, index in tqdm(enumerate(self.metadata)):
paper = self.metadata[index]
title = paper['title'][0]
title_keywords = set() #set(self.parse_doc(title) + self.get_propn(title))
for keyword in set(paper['keyword_search']) | title_keywords:
term = ' '.join(word for word in keyword.lower().split() if word.lower() not in self.stopwords)
if term not in self.index:
self.index[term] = []
self.index[term].append(paper['arxiv_id'])
with open(self.index_path, 'w') as f:
json.dump(self.index, f)
def load_index(self):
print("Loading existing index...")
with open(self.index_path, 'rb') as f:
self.index = json.load(f)
print("Index loaded successfully.")
def load_or_build_index(self):
if os.path.exists(self.index_path):
self.load_index()
else:
self.build_index()
def parse_doc(self, doc):
local_kws = []
for phrase in doc._.phrases:
local_kws.append(phrase.text.lower())
return [self.preprocess_text(word) for word in local_kws]
def get_propn(self, doc):
result = []
working_str = ''
for token in doc:
if(token.text in nlp.Defaults.stop_words or token.text in punctuation):
if working_str != '':
result.append(working_str.strip())
working_str = ''
if(token.pos_ == "PROPN"):
working_str += token.text + ' '
if working_str != '': result.append(working_str.strip())
return [self.preprocess_text(word) for word in result]
def filter(self, query: str, doc_ids = None):
doc = nlp(query)
query_keywords = self.parse_doc(doc)
nouns = self.get_propn(doc)
if self.verbose: print('keywords:', query_keywords)
if self.verbose: print('proper nouns:', nouns)
filtered = set()
if len(query_keywords) > 0 and not self.ne_only:
for keyword in query_keywords:
if keyword != '' and keyword in self.index.keys(): filtered |= set(self.index[keyword])
if len(nouns) > 0:
ne_results = set()
for noun in nouns:
if noun in self.index.keys(): ne_results |= set(self.index[noun])
if self.ne_only: filtered = ne_results # keep only named entity results
else: filtered &= ne_results # take the intersection
if doc_ids is not None: filtered &= doc_ids # apply filter to results
return filtered
def get_cluster_keywords(clust_ids, all_keywords):
tagstr = ''
clust_tags = []
for i in range(len(clust_ids)):
clust_paper_kw = []
for j in range(len(all_keywords[clust_ids[i]])):
clust_tags.append(all_keywords[clust_ids[i]][j])
tags = Counter(clust_tags).most_common(30)
for i in range(len(tags)):
# print(tags[i][0])
if len(tags[i][0]) > 2:
tagstr = tagstr + tags[i][0]+ ', '
return tagstr
def get_keywords(query, ret_indices, all_keywords):
kws = get_cluster_keywords(ret_indices, all_keywords)
kw_prompt = """You are an expert research assistant. Here are a list of keywords corresponding to the topics that a query and its answer are about that you need to synthesize into a succinct summary:
["""+kws+"""]
First, find the keywords that are most relevant to answering the question, and then print them in numbered order. Keywords should be a few words at most. Do not list more than five keywords.
If there are no relevant quotes, write “No relevant keywords” instead.
Thus, the format of your overall response should look like what’s shown between the tags. Make sure to follow the formatting and spacing exactly.
Keywords:
[1] Milky Way galaxy
[2] Good agreement
[3] Bayesian
[4] Observational constraints
[5] Globular clusters
[6] Kinematic data
If the question cannot be answered by the document, say so."""
client = anthropic.Anthropic(api_key=anthropic_key,)
message = client.messages.create(model="claude-3-haiku-20240307",max_tokens=200,temperature=0,system=kw_prompt,
messages=[{"role": "user","content": [{"type": "text","text": query}]}])
return message.content[0].text