Spaces:
Sleeping
Sleeping
linked up everything, qn type, consensus
Browse files- app.py +658 -441
- data/data-00000-of-00001.arrow +3 -0
- data/dataset_info.json +134 -0
- data/state.json +13 -0
- requirements.txt +2 -1
app.py
CHANGED
@@ -6,7 +6,7 @@ from abc import ABC, abstractmethod
|
|
6 |
from typing import List, Dict, Any, Tuple
|
7 |
from collections import defaultdict
|
8 |
from tqdm import tqdm
|
9 |
-
import pandas as pd
|
10 |
from datetime import datetime, date
|
11 |
from datasets import load_dataset, load_from_disk
|
12 |
from collections import Counter
|
@@ -16,24 +16,27 @@ import concurrent.futures
|
|
16 |
|
17 |
from langchain import hub
|
18 |
from langchain_openai import ChatOpenAI as openai_llm
|
19 |
-
from
|
|
|
|
|
20 |
from langchain_community.callbacks import StreamlitCallbackHandler
|
21 |
-
|
22 |
-
from langchain.agents import create_react_agent, Tool, AgentExecutor
|
23 |
from langchain_community.utilities import DuckDuckGoSearchAPIWrapper
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
24 |
|
25 |
-
|
26 |
-
|
27 |
-
|
28 |
-
|
29 |
-
anthropic_key = st.secrets["anthropic_key"]
|
30 |
-
|
31 |
-
openai_key = st.secrets["openai_key"]
|
32 |
|
33 |
from nltk.corpus import stopwords
|
34 |
import nltk
|
35 |
from openai import OpenAI
|
36 |
-
import anthropic
|
37 |
import cohere
|
38 |
import faiss
|
39 |
|
@@ -50,12 +53,28 @@ except:
|
|
50 |
nltk.download('stopwords')
|
51 |
stopwords.words('english')
|
52 |
|
|
|
53 |
from bokeh.plotting import figure
|
54 |
from bokeh.models import ColumnDataSource
|
55 |
from bokeh.io import output_notebook
|
56 |
from bokeh.palettes import Spectral5
|
57 |
from bokeh.transform import linear_cmap
|
58 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
59 |
|
60 |
st.image('local_files/pathfinder_logo.png')
|
61 |
|
@@ -63,15 +82,13 @@ st.expander("About", expanded=False).write(
|
|
63 |
"""
|
64 |
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
|
65 |
sensitivity from modern large language models (LLMs) to better parse patterns in paper contexts.
|
66 |
-
|
67 |
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.
|
68 |
|
69 |
-
**👈
|
70 |
-
of what this framework can do!
|
71 |
|
72 |
### Tool summary:
|
73 |
- Please wait while the initial data loads and compiles, this takes about a minute initially.
|
74 |
-
- `Paper search` looks for relevant papers given an arxiv id or a question.
|
75 |
|
76 |
This is not meant to be a replacement to existing tools like the
|
77 |
[ADS](https://ui.adsabs.harvard.edu/),
|
@@ -79,33 +96,34 @@ st.expander("About", expanded=False).write(
|
|
79 |
that otherwise might be missed during a literature survey.
|
80 |
It is trained on astro-ph (astrophysics of galaxies) papers up to last-year-ish mined from arxiv and supplemented with ADS metadata,
|
81 |
if you are interested in extending it please reach out!
|
82 |
-
|
83 |
-
|
84 |
-
Also add:
|
85 |
|
86 |
The image below shows a representation of all the astro-ph.GA papers that can be explored in more detail
|
87 |
using the `Arxiv embedding` page. The papers tend to cluster together by similarity, and result in an
|
88 |
atlas that shows well studied (forests) and currently uncharted areas (water).
|
89 |
"""
|
90 |
)
|
91 |
-
|
92 |
-
|
93 |
-
|
94 |
-
|
95 |
if 'arxiv_corpus' not in st.session_state:
|
96 |
with st.spinner('loading data...'):
|
97 |
-
try:
|
98 |
arxiv_corpus = load_from_disk('data/')
|
99 |
except:
|
100 |
st.write('downloading data')
|
101 |
-
arxiv_corpus = load_dataset('kiyer/pathfinder_arxiv_data',split='train')
|
|
|
102 |
arxiv_corpus.save_to_disk('data/')
|
103 |
arxiv_corpus.add_faiss_index('embed')
|
104 |
st.session_state.arxiv_corpus = arxiv_corpus
|
105 |
st.toast('loaded arxiv corpus')
|
106 |
else:
|
107 |
arxiv_corpus = st.session_state.arxiv_corpus
|
108 |
-
|
109 |
if 'ids' not in st.session_state:
|
110 |
st.session_state.ids = arxiv_corpus['ads_id']
|
111 |
st.session_state.titles = arxiv_corpus['title']
|
@@ -114,8 +132,8 @@ if 'ids' not in st.session_state:
|
|
114 |
st.session_state.years = arxiv_corpus['date']
|
115 |
st.session_state.kws = arxiv_corpus['keywords']
|
116 |
st.toast('done caching. time taken: %.2f sec' %(time.time()-ts))
|
117 |
-
|
118 |
-
|
119 |
#---------------------------------------------------------------
|
120 |
|
121 |
# A hack to "clear" the previous result when submitting a new prompt. This avoids
|
@@ -144,186 +162,33 @@ def with_clear_container(submit_clicked: bool) -> bool:
|
|
144 |
return True
|
145 |
|
146 |
return False
|
147 |
-
|
148 |
-
#----------------------------------------------------------------
|
149 |
-
|
150 |
-
class Filter():
|
151 |
-
def filter(self, query: str, arxiv_id: str) -> List[str]:
|
152 |
-
pass
|
153 |
-
|
154 |
-
class CitationFilter(Filter): # can do it with all metadata
|
155 |
-
def __init__(self, corpus):
|
156 |
-
self.corpus = corpus
|
157 |
-
ids = ids
|
158 |
-
cites = cites
|
159 |
-
self.citation_counts = {ids[i]: cites[i] for i in range(len(ids))}
|
160 |
-
|
161 |
-
def citation_weight(self, x, shift, scale):
|
162 |
-
return 1 / (1 + np.exp(-1 * (x - shift) / scale)) # sigmoid function
|
163 |
-
|
164 |
-
def filter(self, doc_scores, weight = 0.1): # additive weighting
|
165 |
-
citation_count = np.array([self.citation_counts[doc[0]] for doc in doc_scores])
|
166 |
-
cmean, cstd = np.median(citation_count), np.std(citation_count)
|
167 |
-
citation_score = self.citation_weight(citation_count, cmean, cstd)
|
168 |
-
|
169 |
-
for i, doc in enumerate(doc_scores):
|
170 |
-
doc_scores[i][2] += weight * citation_score[i]
|
171 |
-
|
172 |
-
class DateFilter(Filter): # include time weighting eventually
|
173 |
-
def __init__(self, document_dates):
|
174 |
-
self.document_dates = document_dates
|
175 |
-
|
176 |
-
def parse_date(self, arxiv_id: str) -> datetime: # only for documents
|
177 |
-
if arxiv_id.startswith('astro-ph'):
|
178 |
-
arxiv_id = arxiv_id.split('astro-ph')[1].split('_arXiv')[0]
|
179 |
-
try:
|
180 |
-
year = int("20" + arxiv_id[:2])
|
181 |
-
month = int(arxiv_id[2:4])
|
182 |
-
except:
|
183 |
-
year = 2023
|
184 |
-
month = 1
|
185 |
-
return date(year, month, 1)
|
186 |
-
|
187 |
-
def weight(self, time, shift, scale):
|
188 |
-
return 1 / (1 + np.exp((time - shift) / scale))
|
189 |
-
|
190 |
-
def evaluate_filter(self, year, filter_string):
|
191 |
-
try:
|
192 |
-
# Use ast.literal_eval to safely evaluate the expression
|
193 |
-
result = eval(filter_string, {"__builtins__": None}, {"year": year})
|
194 |
-
return result
|
195 |
-
except Exception as e:
|
196 |
-
print(f"Error evaluating filter: {e}")
|
197 |
-
return False
|
198 |
-
|
199 |
-
def filter(self, docs, boolean_date = None, min_date = None, max_date = None, time_score = 0):
|
200 |
-
filtered = []
|
201 |
-
|
202 |
-
if boolean_date is not None:
|
203 |
-
boolean_date = boolean_date.replace("AND", "and").replace("OR", "or")
|
204 |
-
for doc in docs:
|
205 |
-
if self.evaluate_filter(self.document_dates[doc[0]].year, boolean_date):
|
206 |
-
filtered.append(doc)
|
207 |
-
|
208 |
-
else:
|
209 |
-
if min_date == None: min_date = date(1990, 1, 1)
|
210 |
-
if max_date == None: max_date = date(2024, 7, 3)
|
211 |
-
|
212 |
-
for doc in docs:
|
213 |
-
if self.document_dates[doc[0]] >= min_date and self.document_dates[doc[0]] <= max_date:
|
214 |
-
filtered.append(doc)
|
215 |
-
|
216 |
-
if time_score is not None: # apply time weighting
|
217 |
-
for i, item in enumerate(filtered):
|
218 |
-
time_diff = (max_date - self.document_dates[filtered[i][0]]).days / 365
|
219 |
-
filtered[i][2] += time_score * 0.1 * self.weight(time_diff, 5, 5)
|
220 |
-
|
221 |
-
return filtered
|
222 |
-
|
223 |
-
class KeywordFilter(Filter):
|
224 |
-
def __init__(self, corpus,
|
225 |
-
remove_capitals: bool = True, metadata = None, ne_only = True, verbose = False):
|
226 |
-
|
227 |
-
self.index_path = 'keyword_index.json'
|
228 |
-
# self.metadata = metadata
|
229 |
-
self.remove_capitals = remove_capitals
|
230 |
-
self.ne_only = ne_only
|
231 |
-
self.stopwords = set(stopwords.words('english'))
|
232 |
-
self.verbose = verbose
|
233 |
-
self.index = None
|
234 |
-
self.kws = st.session_state.kws
|
235 |
-
self.ids = st.session_state.ids
|
236 |
-
self.titles = st.session_state.titles
|
237 |
-
|
238 |
-
self.load_or_build_index()
|
239 |
-
|
240 |
-
def preprocess_text(self, text: str) -> str:
|
241 |
-
text = ''.join(char for char in text if char.isalnum() or char.isspace())
|
242 |
-
if self.remove_capitals: text = text.lower()
|
243 |
-
return ' '.join(word for word in text.split() if word.lower() not in self.stopwords)
|
244 |
-
|
245 |
-
def build_index(self): # include the title in the index
|
246 |
-
print("Building index...")
|
247 |
-
self.index = {}
|
248 |
-
|
249 |
-
for i in range(len(self.kws)):
|
250 |
-
paper = self.ids[i]
|
251 |
-
title = self.titles[i]
|
252 |
-
title_keywords = set()
|
253 |
-
for keyword in set(self.kws[i]) | title_keywords:
|
254 |
-
term = ' '.join(word for word in keyword.lower().split() if word.lower() not in self.stopwords)
|
255 |
-
if term not in self.index:
|
256 |
-
self.index[term] = []
|
257 |
-
self.index[term].append(self.ids[i])
|
258 |
-
|
259 |
-
with open(self.index_path, 'w') as f:
|
260 |
-
json.dump(self.index, f)
|
261 |
-
|
262 |
-
def load_index(self):
|
263 |
-
print("Loading existing index...")
|
264 |
-
with open(self.index_path, 'rb') as f:
|
265 |
-
self.index = json.load(f)
|
266 |
-
|
267 |
-
print("Index loaded successfully.")
|
268 |
-
|
269 |
-
def load_or_build_index(self):
|
270 |
-
if os.path.exists(self.index_path):
|
271 |
-
self.load_index()
|
272 |
-
else:
|
273 |
-
self.build_index()
|
274 |
|
275 |
-
|
276 |
-
|
277 |
-
|
278 |
-
|
279 |
-
|
280 |
-
|
281 |
-
|
282 |
-
|
283 |
-
|
284 |
-
|
285 |
-
|
286 |
-
|
287 |
-
|
288 |
-
|
289 |
-
|
290 |
-
|
291 |
-
|
292 |
-
|
293 |
-
|
294 |
-
|
295 |
-
|
296 |
-
if working_str != '': result.append(working_str.strip())
|
297 |
-
|
298 |
-
return [self.preprocess_text(word) for word in result]
|
299 |
-
|
300 |
-
def filter(self, query: str, doc_ids = None):
|
301 |
-
doc = nlp(query)
|
302 |
-
query_keywords = self.parse_doc(doc)
|
303 |
-
nouns = self.get_propn(doc)
|
304 |
-
if self.verbose: print('keywords:', query_keywords)
|
305 |
-
if self.verbose: print('proper nouns:', nouns)
|
306 |
-
|
307 |
-
filtered = set()
|
308 |
-
if len(query_keywords) > 0 and not self.ne_only:
|
309 |
-
for keyword in query_keywords:
|
310 |
-
if keyword != '' and keyword in self.index.keys(): filtered |= set(self.index[keyword])
|
311 |
-
|
312 |
-
if len(nouns) > 0:
|
313 |
-
ne_results = set()
|
314 |
-
for noun in nouns:
|
315 |
-
if noun in self.index.keys(): ne_results |= set(self.index[noun])
|
316 |
-
|
317 |
-
if self.ne_only: filtered = ne_results # keep only named entity results
|
318 |
-
else: filtered &= ne_results # take the intersection
|
319 |
-
|
320 |
-
if doc_ids is not None: filtered &= doc_ids # apply filter to results
|
321 |
-
return filtered
|
322 |
|
323 |
class EmbeddingRetrievalSystem():
|
324 |
|
325 |
def __init__(self, weight_citation = False, weight_date = False, weight_keywords = False):
|
326 |
-
|
327 |
self.ids = st.session_state.ids
|
328 |
self.years = st.session_state.years
|
329 |
self.abstract = st.session_state.abstracts
|
@@ -331,7 +196,8 @@ class EmbeddingRetrievalSystem():
|
|
331 |
self.embed_model = "text-embedding-3-small"
|
332 |
self.dataset = arxiv_corpus
|
333 |
self.kws = st.session_state.kws
|
334 |
-
|
|
|
335 |
self.weight_citation = weight_citation
|
336 |
self.weight_date = weight_date
|
337 |
self.weight_keywords = weight_keywords
|
@@ -339,7 +205,7 @@ class EmbeddingRetrievalSystem():
|
|
339 |
|
340 |
# self.citation_filter = CitationFilter(self.dataset)
|
341 |
# self.date_filter = DateFilter(self.dataset['date'])
|
342 |
-
self.keyword_filter = KeywordFilter(corpus=self.dataset, remove_capitals=True)
|
343 |
|
344 |
def parse_date(self, id):
|
345 |
# indexval = np.where(self.ids == id)[0][0]
|
@@ -354,12 +220,6 @@ class EmbeddingRetrievalSystem():
|
|
354 |
embeddings = self.client.embeddings.create(input=texts, model=self.embed_model).data
|
355 |
return [np.array(embedding.embedding, dtype=np.float32) for embedding in embeddings]
|
356 |
|
357 |
-
def init_filters(self):
|
358 |
-
|
359 |
-
self.citation_filter = []
|
360 |
-
self.date_filter = []
|
361 |
-
self.keyword_filter = []
|
362 |
-
|
363 |
def get_query_embedding(self, query):
|
364 |
return self.make_embedding(query)
|
365 |
|
@@ -370,22 +230,77 @@ class EmbeddingRetrievalSystem():
|
|
370 |
# xq = query_embedding.reshape(-1,1).T.astype('float32')
|
371 |
# D, I = self.index.search(xq, top_k)
|
372 |
# return I[0], D[0]
|
373 |
-
tmp = self.dataset.search('embed',query_embedding, k=top_k)
|
374 |
return [tmp.indices, tmp.scores]
|
375 |
-
|
376 |
def rank_and_filter(self, query, query_embedding, query_date, top_k = 10, return_scores=False, time_result=None):
|
377 |
|
378 |
-
|
379 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
380 |
|
381 |
-
|
382 |
-
|
383 |
-
|
384 |
-
|
385 |
-
|
386 |
-
|
387 |
-
|
388 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
389 |
|
390 |
filtered_results = [[topk_indices[i], similarities[i]] for i in range(len(similarities))]
|
391 |
top_results = sorted(filtered_results, key=lambda x: x[1], reverse=True)[:top_k]
|
@@ -395,43 +310,43 @@ class EmbeddingRetrievalSystem():
|
|
395 |
|
396 |
# Only keep the document IDs
|
397 |
top_results = [doc[0] for doc in top_results]
|
398 |
-
return top_results
|
399 |
-
|
400 |
def retrieve(self, query, top_k, time_result=None, query_date = None, return_scores = False):
|
401 |
|
402 |
query_embedding = self.get_query_embedding(query)
|
403 |
|
404 |
# Judge time relevance
|
405 |
if time_result is None:
|
406 |
-
if self.weight_date:
|
407 |
time_result, time_taken = self.analyze_temporal_query(query, self.anthropic_client)
|
408 |
-
else:
|
409 |
time_result = {'has_temporal_aspect': False, 'expected_year_filter': None, 'expected_recency_weight': None}
|
410 |
|
411 |
-
top_results = self.rank_and_filter(query,
|
412 |
-
query_embedding,
|
413 |
-
query_date,
|
414 |
-
top_k,
|
415 |
-
return_scores = return_scores,
|
416 |
time_result = time_result)
|
417 |
-
|
418 |
return top_results
|
419 |
|
420 |
class HydeRetrievalSystem(EmbeddingRetrievalSystem):
|
421 |
-
def __init__(self, generation_model: str = "claude-3-haiku-20240307",
|
422 |
-
embedding_model: str = "text-embedding-3-small",
|
423 |
-
temperature: float = 0.5,
|
424 |
-
max_doclen: int = 500,
|
425 |
-
generate_n: int = 1,
|
426 |
-
embed_query = True,
|
427 |
conclusion = False, **kwargs):
|
428 |
-
|
429 |
# Handle the kwargs for the superclass init -- filters/citation weighting
|
430 |
super().__init__(**kwargs)
|
431 |
-
|
432 |
if max_doclen * generate_n > 8191:
|
433 |
raise ValueError("Too many tokens. Please reduce max_doclen or generate_n.")
|
434 |
-
|
435 |
self.embedding_model = embedding_model
|
436 |
self.generation_model = generation_model
|
437 |
|
@@ -442,58 +357,67 @@ class HydeRetrievalSystem(EmbeddingRetrievalSystem):
|
|
442 |
self.embed_query = embed_query # embed the query vector?
|
443 |
self.conclusion = conclusion # generate conclusion as well?
|
444 |
|
445 |
-
self.anthropic_key = anthropic_key
|
446 |
-
self.generation_client = anthropic.Anthropic(api_key = self.anthropic_key)
|
447 |
-
|
|
|
448 |
def retrieve(self, query: str, top_k: int = 10, return_scores = False, time_result = None) -> List[Tuple[str, str, float]]:
|
449 |
if time_result is None:
|
450 |
if self.weight_date: time_result, time_taken = analyze_temporal_query(query, self.anthropic_client)
|
451 |
else: time_result = {'has_temporal_aspect': False, 'expected_year_filter': None, 'expected_recency_weight': None}
|
452 |
|
453 |
docs = self.generate_docs(query)
|
|
|
|
|
454 |
doc_embeddings = self.embed_docs(docs)
|
455 |
|
456 |
-
if self.embed_query:
|
457 |
query_emb = self.embed_docs([query])[0]
|
458 |
doc_embeddings.append(query_emb)
|
459 |
-
|
460 |
embedding = np.mean(np.array(doc_embeddings), axis = 0)
|
461 |
|
462 |
top_results = self.rank_and_filter(query, embedding, query_date=None, top_k = top_k, return_scores = return_scores, time_result = time_result)
|
463 |
-
|
464 |
return top_results
|
465 |
|
466 |
def generate_doc(self, query: str):
|
467 |
-
prompt = """You are an expert astronomer. Given a scientific query, generate the abstract
|
468 |
-
if self.conclusion:
|
469 |
-
prompt += " and conclusion"
|
470 |
-
prompt += """ of an expert-level research paper
|
471 |
that answers the question. Stick to a maximum length of {} tokens and return just the text of the abstract and conclusion.
|
472 |
Do not include labels for any section. Use research-specific jargon.""".format(self.max_doclen)
|
473 |
-
|
474 |
-
|
475 |
-
message = self.generation_client.messages.create(
|
476 |
-
|
477 |
-
|
478 |
-
|
479 |
-
|
480 |
-
|
481 |
-
|
482 |
-
|
483 |
-
|
484 |
-
|
485 |
-
|
|
|
|
|
|
|
|
|
486 |
def generate_docs(self, query: str):
|
487 |
docs = []
|
488 |
-
|
489 |
-
|
490 |
-
|
491 |
-
|
492 |
-
|
493 |
-
|
494 |
-
|
495 |
-
|
496 |
-
|
|
|
|
|
|
|
|
|
|
|
497 |
return docs
|
498 |
|
499 |
def embed_docs(self, docs: List[str]):
|
@@ -503,35 +427,35 @@ class HydeCohereRetrievalSystem(HydeRetrievalSystem):
|
|
503 |
def __init__(self, **kwargs):
|
504 |
super().__init__(**kwargs)
|
505 |
|
506 |
-
self.cohere_key =
|
507 |
self.cohere_client = cohere.Client(self.cohere_key)
|
508 |
|
509 |
-
def retrieve(self, query: str,
|
510 |
-
top_k: int = 10,
|
511 |
rerank_top_k: int = 250,
|
512 |
return_scores = False, time_result = None,
|
513 |
reweight = False) -> List[Tuple[str, str, float]]:
|
514 |
-
|
515 |
if time_result is None:
|
516 |
if self.weight_date: time_result, time_taken = analyze_temporal_query(query, self.anthropic_client)
|
517 |
else: time_result = {'has_temporal_aspect': False, 'expected_year_filter': None, 'expected_recency_weight': None}
|
518 |
-
|
519 |
top_results = super().retrieve(query, top_k = rerank_top_k, time_result = time_result)
|
520 |
-
|
521 |
# doc_texts = self.get_document_texts(top_results)
|
522 |
# docs_for_rerank = [f"Abstract: {doc['abstract']}\nConclusions: {doc['conclusions']}" for doc in doc_texts]
|
523 |
docs_for_rerank = [self.abstract[i] for i in top_results]
|
524 |
-
|
525 |
if len(docs_for_rerank) == 0:
|
526 |
return []
|
527 |
-
|
528 |
reranked_results = self.cohere_client.rerank(
|
529 |
query=query,
|
530 |
documents=docs_for_rerank,
|
531 |
model='rerank-english-v3.0',
|
532 |
top_n=top_k
|
533 |
)
|
534 |
-
|
535 |
final_results = []
|
536 |
for result in reranked_results.results:
|
537 |
doc_id = top_results[result.index]
|
@@ -542,9 +466,9 @@ class HydeCohereRetrievalSystem(HydeRetrievalSystem):
|
|
542 |
if reweight:
|
543 |
if time_result['has_temporal_aspect']:
|
544 |
final_results = self.date_filter.filter(final_results, time_score = time_result['expected_recency_weight'])
|
545 |
-
|
546 |
if self.weight_citation: self.citation_filter.filter(final_results)
|
547 |
-
|
548 |
if return_scores:
|
549 |
return {result[0]: result[2] for result in final_results}
|
550 |
|
@@ -554,40 +478,113 @@ class HydeCohereRetrievalSystem(HydeRetrievalSystem):
|
|
554 |
return self.embed_batch(docs)
|
555 |
|
556 |
# ----------------------------------------------------------------
|
557 |
-
|
558 |
-
|
559 |
if 'ec' not in st.session_state:
|
560 |
-
ec =
|
561 |
st.session_state.ec = ec
|
562 |
st.toast('loaded retrieval system')
|
563 |
else:
|
564 |
ec = st.session_state.ec
|
565 |
-
|
566 |
-
# Function to simulate question answering (replace with actual implementation)
|
567 |
-
def answer_question(question, top_k, keywords, toggles, method, question_type):
|
568 |
-
# Simulated answer (replace with actual logic)
|
569 |
-
# return f"Answer to '{question}' using method {method} for {question_type} question."
|
570 |
-
return run_ret(question, top_k)
|
571 |
|
|
|
|
|
|
|
|
|
572 |
|
573 |
-
def
|
574 |
-
|
575 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
576 |
for i in ids:
|
577 |
papers.append(st.session_state.titles[i])
|
578 |
scores.append(ids[i])
|
579 |
links.append('https://ui.adsabs.harvard.edu/abs/'+st.session_state.arxiv_corpus['bibcode'][i]+'/abstract')
|
580 |
-
|
|
|
|
|
|
|
581 |
return pd.DataFrame({
|
582 |
'Title': papers,
|
583 |
'Relevance': scores,
|
584 |
-
'
|
|
|
|
|
|
|
585 |
})
|
586 |
|
587 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
588 |
|
589 |
def create_embedding_plot(rs):
|
590 |
-
|
|
|
|
|
591 |
|
592 |
pltsource = ColumnDataSource(data=dict(
|
593 |
x=st.session_state.arxiv_corpus['umap_x'],
|
@@ -595,10 +592,14 @@ def create_embedding_plot(rs):
|
|
595 |
title=st.session_state.titles,
|
596 |
link=st.session_state.arxiv_corpus['bibcode'],
|
597 |
))
|
598 |
-
|
599 |
rsflag = np.zeros((len(st.session_state.ids),))
|
600 |
rsflag[np.array([k for k in rs])] = 1
|
|
|
|
|
|
|
601 |
pltsource.data['colors'] = rsflag * 0.8 + 0.1
|
|
|
602 |
pltsource.data['sizes'] = (rsflag + 1)**5 / 100
|
603 |
|
604 |
TOOLTIPS = """
|
@@ -609,22 +610,21 @@ def create_embedding_plot(rs):
|
|
609 |
@link <br> <br>
|
610 |
</div>
|
611 |
"""
|
612 |
-
|
613 |
mapper = linear_cmap(field_name="colors", palette=Spectral5, low=0., high=1.)
|
614 |
|
615 |
p = figure(width=700, height=900, tooltips=TOOLTIPS, x_range=(0, 20), y_range=(-4.2,18),
|
616 |
title="UMAP projection of embeddings for the astro-ph corpus")
|
617 |
-
|
618 |
p.axis.visible=False
|
619 |
p.grid.visible=False
|
620 |
p.outline_line_alpha = 0.
|
621 |
-
|
622 |
p.circle('x', 'y', radius='sizes', source=pltsource, alpha=0.3, fill_color=mapper, fill_alpha='colors', line_color="lightgrey",line_alpha=0.1)
|
623 |
-
|
624 |
return p
|
625 |
|
626 |
-
|
627 |
-
def extract_keywords(question):
|
628 |
# Simulated keyword extraction (replace with actual logic)
|
629 |
return ['keyword1', 'keyword2', 'keyword3']
|
630 |
|
@@ -633,184 +633,401 @@ def estimate_consensus():
|
|
633 |
# Simulated consensus estimation (replace with actual calculation)
|
634 |
return 0.75
|
635 |
|
636 |
-
def run_ret(query, top_k):
|
637 |
-
rs = ec.retrieve(query, top_k, return_scores=True)
|
638 |
-
output_str = ''
|
639 |
-
for i in rs:
|
640 |
-
if rs[i] > 0.5:
|
641 |
-
output_str = output_str + '---> ' + st.session_state.abstracts[i] + '(score: %.2f) \n' %rs[i]
|
642 |
-
else:
|
643 |
-
output_str = output_str + st.session_state.abstracts[i] + '(score: %.2f) \n' %rs[i]
|
644 |
-
return output_str, rs
|
645 |
|
646 |
-
def
|
647 |
-
|
648 |
-
|
649 |
-
|
650 |
-
|
651 |
-
|
652 |
-
|
653 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
654 |
|
655 |
-
|
656 |
-
tools =
|
657 |
-
Tool(
|
658 |
-
name="Library",
|
659 |
-
func=Library,
|
660 |
-
description="A source of information pertinent to your question. Do not answer a question without consulting this!"
|
661 |
-
),
|
662 |
-
Tool(
|
663 |
-
name="Search",
|
664 |
-
func=search.run,
|
665 |
-
description="useful for when you need to look up knowledge about common topics or current events",
|
666 |
-
)
|
667 |
-
]
|
668 |
-
|
669 |
-
if 'tools' not in st.session_state:
|
670 |
-
st.session_state.tools = tools
|
671 |
-
|
672 |
-
# for another question type:
|
673 |
-
# First, find the quotes from the document that are most relevant to answering the question, and then print them in numbered order.
|
674 |
-
# Quotes should be relatively short. If there are no relevant quotes, write “No relevant quotes” instead.
|
675 |
|
676 |
-
|
677 |
|
678 |
-
|
679 |
-
|
680 |
-
|
681 |
-
Try to break the question down into smaller steps and solve it in a logical manner.
|
682 |
|
683 |
-
You have access to the following tools:
|
684 |
|
685 |
-
|
|
|
|
|
|
|
686 |
|
687 |
-
|
688 |
|
689 |
-
|
690 |
-
Thought: you should always think about what to do
|
691 |
-
Action: the action to take, should be one of [{tool_names}]
|
692 |
-
Action Input: the input to the action
|
693 |
-
Observation: the result of the action
|
694 |
-
... (this Thought/Action/Action Input/Observation can repeat N times)
|
695 |
-
Thought: I now know the final answer
|
696 |
-
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
|
697 |
|
698 |
-
|
699 |
|
700 |
-
Question:
|
701 |
-
Thought:
|
|
|
|
|
|
|
|
|
|
|
|
|
702 |
|
|
|
703 |
|
704 |
-
|
705 |
-
|
706 |
|
707 |
-
|
708 |
-
|
709 |
-
|
710 |
-
#
|
711 |
-
|
712 |
-
|
713 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
714 |
|
715 |
-
|
716 |
-
|
717 |
-
|
718 |
-
agent = create_react_agent(llm=gen_llm, tools=tools, prompt=prompt)
|
719 |
-
st.session_state.agent = agent
|
720 |
|
721 |
-
|
722 |
-
|
723 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
724 |
|
725 |
|
726 |
# Streamlit app
|
727 |
def main():
|
728 |
-
|
729 |
# st.title("Question Answering App")
|
730 |
-
|
731 |
-
|
732 |
# Sidebar (Inputs)
|
733 |
st.sidebar.header("Fine-tune the search")
|
734 |
top_k = st.sidebar.slider("Number of papers to retrieve:", 3, 30, 10)
|
735 |
extra_keywords = st.sidebar.text_input("Enter extra keywords (comma-separated):")
|
736 |
-
|
737 |
st.sidebar.subheader("Toggles")
|
738 |
-
toggle_a = st.sidebar.
|
739 |
-
toggle_b = st.sidebar.
|
740 |
-
toggle_c = st.sidebar.
|
741 |
-
|
742 |
-
method = st.sidebar.radio("
|
743 |
-
|
744 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
745 |
|
746 |
-
|
747 |
store_output = st.sidebar.button("Save output")
|
748 |
|
749 |
# Main page (Outputs)
|
750 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
751 |
query = st.text_input("Ask me anything:")
|
752 |
submit_button = st.button("Submit")
|
753 |
-
|
754 |
if submit_button:
|
755 |
-
|
756 |
-
|
757 |
-
|
758 |
-
|
759 |
-
|
760 |
-
|
761 |
-
|
762 |
-
|
763 |
-
|
764 |
-
|
765 |
-
|
766 |
-
|
767 |
-
|
768 |
-
|
769 |
-
|
770 |
-
|
771 |
-
|
772 |
-
|
773 |
-
|
774 |
-
|
775 |
-
|
776 |
-
|
777 |
-
|
778 |
-
|
779 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
780 |
with st.expander("Relevant papers", expanded=True):
|
781 |
# st.dataframe(papers_df, hide_index=True)
|
782 |
st.data_editor(papers_df,
|
783 |
-
column_config = {'Link':st.column_config.LinkColumn(display_text= 'https://ui.adsabs.harvard.edu/abs/(.*?)/abstract')}
|
784 |
)
|
785 |
|
786 |
with st.expander("Embedding map", expanded=False):
|
787 |
st.bokeh_chart(embedding_plot)
|
788 |
-
|
789 |
col1, col2 = st.columns(2)
|
790 |
-
|
791 |
with col1:
|
792 |
-
|
793 |
-
st.subheader("Question
|
794 |
-
|
795 |
-
|
796 |
-
|
797 |
-
|
798 |
-
|
|
|
|
|
|
|
799 |
with col2:
|
800 |
-
|
801 |
-
st.subheader("
|
802 |
-
st.write(
|
803 |
-
|
804 |
-
|
805 |
-
|
806 |
-
|
807 |
-
|
808 |
-
|
|
|
|
|
809 |
else:
|
810 |
-
st.info("Use the sidebar to
|
811 |
-
|
812 |
if store_output:
|
813 |
st.toast("Output stored successfully!")
|
814 |
|
815 |
if __name__ == "__main__":
|
816 |
-
main()
|
|
|
6 |
from typing import List, Dict, Any, Tuple
|
7 |
from collections import defaultdict
|
8 |
from tqdm import tqdm
|
9 |
+
import pandas as pd
|
10 |
from datetime import datetime, date
|
11 |
from datasets import load_dataset, load_from_disk
|
12 |
from collections import Counter
|
|
|
16 |
|
17 |
from langchain import hub
|
18 |
from langchain_openai import ChatOpenAI as openai_llm
|
19 |
+
from langchain_openai import OpenAIEmbeddings
|
20 |
+
from langchain_core.runnables import RunnableConfig, RunnablePassthrough, RunnableParallel
|
21 |
+
from langchain_core.prompts import PromptTemplate
|
22 |
from langchain_community.callbacks import StreamlitCallbackHandler
|
|
|
|
|
23 |
from langchain_community.utilities import DuckDuckGoSearchAPIWrapper
|
24 |
+
from langchain_community.vectorstores import Chroma
|
25 |
+
from langchain_community.document_loaders import TextLoader
|
26 |
+
from langchain.agents import create_react_agent, Tool, AgentExecutor
|
27 |
+
from langchain.text_splitter import RecursiveCharacterTextSplitter
|
28 |
+
from langchain_core.output_parsers import StrOutputParser
|
29 |
+
from langchain.callbacks import FileCallbackHandler
|
30 |
+
from langchain.callbacks.manager import CallbackManager
|
31 |
|
32 |
+
import instructor
|
33 |
+
from pydantic import BaseModel, Field
|
34 |
+
from typing import List, Literal
|
|
|
|
|
|
|
|
|
35 |
|
36 |
from nltk.corpus import stopwords
|
37 |
import nltk
|
38 |
from openai import OpenAI
|
39 |
+
# import anthropic
|
40 |
import cohere
|
41 |
import faiss
|
42 |
|
|
|
53 |
nltk.download('stopwords')
|
54 |
stopwords.words('english')
|
55 |
|
56 |
+
|
57 |
from bokeh.plotting import figure
|
58 |
from bokeh.models import ColumnDataSource
|
59 |
from bokeh.io import output_notebook
|
60 |
from bokeh.palettes import Spectral5
|
61 |
from bokeh.transform import linear_cmap
|
62 |
|
63 |
+
ts = time.time()
|
64 |
+
|
65 |
+
|
66 |
+
# anthropic_key = st.secrets["anthropic_key"]
|
67 |
+
openai_key = st.secrets["openai_key"]
|
68 |
+
cohere_key = st.secrets['cohere_key']
|
69 |
+
|
70 |
+
gen_llm = openai_llm(temperature=0,model_name='gpt-4o-mini', openai_api_key = openai_key)
|
71 |
+
consensus_client = instructor.patch(OpenAI(api_key=openai_key))
|
72 |
+
|
73 |
+
embed_client = OpenAI(api_key = openai_key)
|
74 |
+
embed_model = "text-embedding-3-small"
|
75 |
+
embeddings = OpenAIEmbeddings(model = embed_model, api_key = openai_key)
|
76 |
+
|
77 |
+
|
78 |
|
79 |
st.image('local_files/pathfinder_logo.png')
|
80 |
|
|
|
82 |
"""
|
83 |
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
|
84 |
sensitivity from modern large language models (LLMs) to better parse patterns in paper contexts.
|
85 |
+
|
86 |
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.
|
87 |
|
88 |
+
**👈 Use the sidebar to tweak the search parameters to get better results**.
|
|
|
89 |
|
90 |
### Tool summary:
|
91 |
- Please wait while the initial data loads and compiles, this takes about a minute initially.
|
|
|
92 |
|
93 |
This is not meant to be a replacement to existing tools like the
|
94 |
[ADS](https://ui.adsabs.harvard.edu/),
|
|
|
96 |
that otherwise might be missed during a literature survey.
|
97 |
It is trained on astro-ph (astrophysics of galaxies) papers up to last-year-ish mined from arxiv and supplemented with ADS metadata,
|
98 |
if you are interested in extending it please reach out!
|
99 |
+
|
100 |
+
|
101 |
+
Also add: feedback form, socials, literature, contact us, copyright, collaboration, etc.
|
102 |
|
103 |
The image below shows a representation of all the astro-ph.GA papers that can be explored in more detail
|
104 |
using the `Arxiv embedding` page. The papers tend to cluster together by similarity, and result in an
|
105 |
atlas that shows well studied (forests) and currently uncharted areas (water).
|
106 |
"""
|
107 |
)
|
108 |
+
|
109 |
+
|
110 |
+
# ---------------- get data and set up session state ---------------------------
|
111 |
+
|
112 |
if 'arxiv_corpus' not in st.session_state:
|
113 |
with st.spinner('loading data...'):
|
114 |
+
try:
|
115 |
arxiv_corpus = load_from_disk('data/')
|
116 |
except:
|
117 |
st.write('downloading data')
|
118 |
+
# arxiv_corpus = load_dataset('kiyer/pathfinder_arxiv_data',split='train')
|
119 |
+
arxiv_corpus = load_dataset('kiyer/pathfinder_arxiv_data_galaxy',split='train')
|
120 |
arxiv_corpus.save_to_disk('data/')
|
121 |
arxiv_corpus.add_faiss_index('embed')
|
122 |
st.session_state.arxiv_corpus = arxiv_corpus
|
123 |
st.toast('loaded arxiv corpus')
|
124 |
else:
|
125 |
arxiv_corpus = st.session_state.arxiv_corpus
|
126 |
+
|
127 |
if 'ids' not in st.session_state:
|
128 |
st.session_state.ids = arxiv_corpus['ads_id']
|
129 |
st.session_state.titles = arxiv_corpus['title']
|
|
|
132 |
st.session_state.years = arxiv_corpus['date']
|
133 |
st.session_state.kws = arxiv_corpus['keywords']
|
134 |
st.toast('done caching. time taken: %.2f sec' %(time.time()-ts))
|
135 |
+
|
136 |
+
|
137 |
#---------------------------------------------------------------
|
138 |
|
139 |
# A hack to "clear" the previous result when submitting a new prompt. This avoids
|
|
|
162 |
return True
|
163 |
|
164 |
return False
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
165 |
|
166 |
+
# ---------------- define embedding retrieval systems --------------------------
|
167 |
+
|
168 |
+
def get_keywords(text):
|
169 |
+
result = []
|
170 |
+
pos_tag = ['PROPN', 'ADJ', 'NOUN']
|
171 |
+
doc = nlp(text.lower())
|
172 |
+
for token in doc:
|
173 |
+
if(token.text in nlp.Defaults.stop_words or token.text in punctuation):
|
174 |
+
continue
|
175 |
+
if(token.pos_ in pos_tag):
|
176 |
+
result.append(token.text)
|
177 |
+
return result
|
178 |
+
|
179 |
+
def parse_doc(text, nret = 10):
|
180 |
+
local_kws = []
|
181 |
+
doc = nlp(text)
|
182 |
+
# examine the top-ranked phrases in the document
|
183 |
+
for phrase in doc._.phrases[:nret]:
|
184 |
+
# print(phrase.text)
|
185 |
+
local_kws.append(phrase.text)
|
186 |
+
return local_kws
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
187 |
|
188 |
class EmbeddingRetrievalSystem():
|
189 |
|
190 |
def __init__(self, weight_citation = False, weight_date = False, weight_keywords = False):
|
191 |
+
|
192 |
self.ids = st.session_state.ids
|
193 |
self.years = st.session_state.years
|
194 |
self.abstract = st.session_state.abstracts
|
|
|
196 |
self.embed_model = "text-embedding-3-small"
|
197 |
self.dataset = arxiv_corpus
|
198 |
self.kws = st.session_state.kws
|
199 |
+
self.cites = st.session_state.cites
|
200 |
+
|
201 |
self.weight_citation = weight_citation
|
202 |
self.weight_date = weight_date
|
203 |
self.weight_keywords = weight_keywords
|
|
|
205 |
|
206 |
# self.citation_filter = CitationFilter(self.dataset)
|
207 |
# self.date_filter = DateFilter(self.dataset['date'])
|
208 |
+
# self.keyword_filter = KeywordFilter(corpus=self.dataset, remove_capitals=True)
|
209 |
|
210 |
def parse_date(self, id):
|
211 |
# indexval = np.where(self.ids == id)[0][0]
|
|
|
220 |
embeddings = self.client.embeddings.create(input=texts, model=self.embed_model).data
|
221 |
return [np.array(embedding.embedding, dtype=np.float32) for embedding in embeddings]
|
222 |
|
|
|
|
|
|
|
|
|
|
|
|
|
223 |
def get_query_embedding(self, query):
|
224 |
return self.make_embedding(query)
|
225 |
|
|
|
230 |
# xq = query_embedding.reshape(-1,1).T.astype('float32')
|
231 |
# D, I = self.index.search(xq, top_k)
|
232 |
# return I[0], D[0]
|
233 |
+
tmp = self.dataset.search('embed', query_embedding, k=top_k)
|
234 |
return [tmp.indices, tmp.scores]
|
235 |
+
|
236 |
def rank_and_filter(self, query, query_embedding, query_date, top_k = 10, return_scores=False, time_result=None):
|
237 |
|
238 |
+
# st.write('status')
|
239 |
+
|
240 |
+
# st.write('toggles', self.toggles)
|
241 |
+
# st.write('question_type', self.question_type)
|
242 |
+
# st.write('rag method', self.rag_method)
|
243 |
+
# st.write('gen method', self.gen_method)
|
244 |
+
|
245 |
+
self.weight_keywords = self.toggles["Keyword weighting"]
|
246 |
+
self.weight_date = self.toggles["Time weighting"]
|
247 |
+
self.weight_citation = self.toggles["Citation weighting"]
|
248 |
|
249 |
+
topk_indices, similarities = self.calc_faiss(np.array(query_embedding), top_k = 1000)
|
250 |
+
similarities = 1/similarities # converting from a distance (less is better) to a similarity (more is better)
|
251 |
+
|
252 |
+
query_kws = get_keywords(query)
|
253 |
+
input_kws = self.query_input_keywords
|
254 |
+
query_kws = query_kws + input_kws
|
255 |
+
self.query_kws = query_kws
|
256 |
+
|
257 |
+
if self.weight_keywords == True:
|
258 |
+
sub_kws = [self.kws[i] for i in topk_indices]
|
259 |
+
kw_weight = np.zeros((len(topk_indices),)) + 0.1
|
260 |
+
|
261 |
+
for k in query_kws:
|
262 |
+
for i in (range(len(topk_indices))):
|
263 |
+
for j in range(len(sub_kws[i])):
|
264 |
+
if k.lower() in sub_kws[i][j].lower():
|
265 |
+
kw_weight[i] = kw_weight[i] + 0.1
|
266 |
+
# print(i, k, sub_kws[i][j])
|
267 |
+
|
268 |
+
# kw_weight = kw_weight**0.36 / np.amax(kw_weight**0.36)
|
269 |
+
kw_weight = kw_weight / np.amax(kw_weight)
|
270 |
+
else:
|
271 |
+
kw_weight = np.ones((len(topk_indices),))
|
272 |
+
|
273 |
+
if self.weight_date == True:
|
274 |
+
sub_dates = [self.years[i] for i in topk_indices]
|
275 |
+
date = datetime.now().date()
|
276 |
+
date_diff = np.array([((date - i).days / 365.) for i in sub_dates])
|
277 |
+
# age_weight = (1 + np.exp(date_diff/2.1))**(-1) + 0.5
|
278 |
+
age_weight = (1 + np.exp(date_diff/0.7))**(-1)
|
279 |
+
age_weight = age_weight / np.amax(age_weight)
|
280 |
+
else:
|
281 |
+
age_weight = np.ones((len(topk_indices),))
|
282 |
+
|
283 |
+
if self.weight_citation == True:
|
284 |
+
# st.write('weighting by citations')
|
285 |
+
sub_cites = np.array([self.cites[i] for i in topk_indices])
|
286 |
+
temp = sub_cites.copy()
|
287 |
+
temp[sub_cites > 300] = 300.
|
288 |
+
cite_weight = (1 + np.exp((300-temp)/42.0))**(-1.)
|
289 |
+
cite_weight = cite_weight / np.amax(cite_weight)
|
290 |
+
else:
|
291 |
+
cite_weight = np.ones((len(topk_indices),))
|
292 |
+
|
293 |
+
similarities = similarities * (kw_weight) * (age_weight) * (cite_weight)
|
294 |
+
|
295 |
+
# if self.weight_keywords:
|
296 |
+
# keyword_matches = self.keyword_filter.filter(query)
|
297 |
+
# self.query_kws = keyword_matches
|
298 |
+
# kw_indices = np.zeros_like(similarities)
|
299 |
+
# for s in keyword_matches:
|
300 |
+
# if self.id_to_index[s] in topk_indices:
|
301 |
+
# # print('yes', self.id_to_index[s], topk_indices[np.where(topk_indices == self.id_to_index[s])[0]])
|
302 |
+
# similarities[np.where(topk_indices == self.id_to_index[s])[0]] = similarities[np.where(topk_indices == self.id_to_index[s])[0]] * 10.
|
303 |
+
# similarities = similarities / 10.
|
304 |
|
305 |
filtered_results = [[topk_indices[i], similarities[i]] for i in range(len(similarities))]
|
306 |
top_results = sorted(filtered_results, key=lambda x: x[1], reverse=True)[:top_k]
|
|
|
310 |
|
311 |
# Only keep the document IDs
|
312 |
top_results = [doc[0] for doc in top_results]
|
313 |
+
return top_results
|
314 |
+
|
315 |
def retrieve(self, query, top_k, time_result=None, query_date = None, return_scores = False):
|
316 |
|
317 |
query_embedding = self.get_query_embedding(query)
|
318 |
|
319 |
# Judge time relevance
|
320 |
if time_result is None:
|
321 |
+
if self.weight_date:
|
322 |
time_result, time_taken = self.analyze_temporal_query(query, self.anthropic_client)
|
323 |
+
else:
|
324 |
time_result = {'has_temporal_aspect': False, 'expected_year_filter': None, 'expected_recency_weight': None}
|
325 |
|
326 |
+
top_results = self.rank_and_filter(query,
|
327 |
+
query_embedding,
|
328 |
+
query_date,
|
329 |
+
top_k,
|
330 |
+
return_scores = return_scores,
|
331 |
time_result = time_result)
|
332 |
+
|
333 |
return top_results
|
334 |
|
335 |
class HydeRetrievalSystem(EmbeddingRetrievalSystem):
|
336 |
+
def __init__(self, generation_model: str = "claude-3-haiku-20240307",
|
337 |
+
embedding_model: str = "text-embedding-3-small",
|
338 |
+
temperature: float = 0.5,
|
339 |
+
max_doclen: int = 500,
|
340 |
+
generate_n: int = 1,
|
341 |
+
embed_query = True,
|
342 |
conclusion = False, **kwargs):
|
343 |
+
|
344 |
# Handle the kwargs for the superclass init -- filters/citation weighting
|
345 |
super().__init__(**kwargs)
|
346 |
+
|
347 |
if max_doclen * generate_n > 8191:
|
348 |
raise ValueError("Too many tokens. Please reduce max_doclen or generate_n.")
|
349 |
+
|
350 |
self.embedding_model = embedding_model
|
351 |
self.generation_model = generation_model
|
352 |
|
|
|
357 |
self.embed_query = embed_query # embed the query vector?
|
358 |
self.conclusion = conclusion # generate conclusion as well?
|
359 |
|
360 |
+
# self.anthropic_key = anthropic_key
|
361 |
+
# self.generation_client = anthropic.Anthropic(api_key = self.anthropic_key)
|
362 |
+
self.generation_client = openai_llm(temperature=0,model_name='gpt-4o-mini', openai_api_key = openai_key)
|
363 |
+
|
364 |
def retrieve(self, query: str, top_k: int = 10, return_scores = False, time_result = None) -> List[Tuple[str, str, float]]:
|
365 |
if time_result is None:
|
366 |
if self.weight_date: time_result, time_taken = analyze_temporal_query(query, self.anthropic_client)
|
367 |
else: time_result = {'has_temporal_aspect': False, 'expected_year_filter': None, 'expected_recency_weight': None}
|
368 |
|
369 |
docs = self.generate_docs(query)
|
370 |
+
st.expander('Abstract generated with hyde', expanded=False).write(docs)
|
371 |
+
|
372 |
doc_embeddings = self.embed_docs(docs)
|
373 |
|
374 |
+
if self.embed_query:
|
375 |
query_emb = self.embed_docs([query])[0]
|
376 |
doc_embeddings.append(query_emb)
|
377 |
+
|
378 |
embedding = np.mean(np.array(doc_embeddings), axis = 0)
|
379 |
|
380 |
top_results = self.rank_and_filter(query, embedding, query_date=None, top_k = top_k, return_scores = return_scores, time_result = time_result)
|
381 |
+
|
382 |
return top_results
|
383 |
|
384 |
def generate_doc(self, query: str):
|
385 |
+
prompt = """You are an expert astronomer. Given a scientific query, generate the abstract of an expert-level research paper
|
|
|
|
|
|
|
386 |
that answers the question. Stick to a maximum length of {} tokens and return just the text of the abstract and conclusion.
|
387 |
Do not include labels for any section. Use research-specific jargon.""".format(self.max_doclen)
|
388 |
+
# st.write('invoking hyde generation')
|
389 |
+
|
390 |
+
# message = self.generation_client.messages.create(
|
391 |
+
# model = self.generation_model,
|
392 |
+
# max_tokens = self.max_doclen,
|
393 |
+
# temperature = self.temperature,
|
394 |
+
# system = prompt,
|
395 |
+
# messages=[{ "role": "user",
|
396 |
+
# "content": [{"type": "text", "text": query,}] }]
|
397 |
+
# )
|
398 |
+
# return message.content[0].text
|
399 |
+
|
400 |
+
messages = [("system",prompt,),("human", query),]
|
401 |
+
return self.generation_client.invoke(messages).content
|
402 |
+
|
403 |
+
|
404 |
+
|
405 |
def generate_docs(self, query: str):
|
406 |
docs = []
|
407 |
+
for i in range(self.generate_n):
|
408 |
+
# st.write('invoking hyde generation2')
|
409 |
+
|
410 |
+
docs.append(self.generate_doc(query))
|
411 |
+
# with concurrent.futures.ThreadPoolExecutor() as executor:
|
412 |
+
# st.write('invoking hyde generation2')
|
413 |
+
# future_to_query = {executor.submit(self.generate_doc, query): query for i in range(self.generate_n)}
|
414 |
+
# for future in concurrent.futures.as_completed(future_to_query):
|
415 |
+
# query = future_to_query[future]
|
416 |
+
# try:
|
417 |
+
# data = future.result()
|
418 |
+
# docs.append(data)
|
419 |
+
# except Exception as exc:
|
420 |
+
# pass
|
421 |
return docs
|
422 |
|
423 |
def embed_docs(self, docs: List[str]):
|
|
|
427 |
def __init__(self, **kwargs):
|
428 |
super().__init__(**kwargs)
|
429 |
|
430 |
+
self.cohere_key = cohere_key
|
431 |
self.cohere_client = cohere.Client(self.cohere_key)
|
432 |
|
433 |
+
def retrieve(self, query: str,
|
434 |
+
top_k: int = 10,
|
435 |
rerank_top_k: int = 250,
|
436 |
return_scores = False, time_result = None,
|
437 |
reweight = False) -> List[Tuple[str, str, float]]:
|
438 |
+
|
439 |
if time_result is None:
|
440 |
if self.weight_date: time_result, time_taken = analyze_temporal_query(query, self.anthropic_client)
|
441 |
else: time_result = {'has_temporal_aspect': False, 'expected_year_filter': None, 'expected_recency_weight': None}
|
442 |
+
|
443 |
top_results = super().retrieve(query, top_k = rerank_top_k, time_result = time_result)
|
444 |
+
|
445 |
# doc_texts = self.get_document_texts(top_results)
|
446 |
# docs_for_rerank = [f"Abstract: {doc['abstract']}\nConclusions: {doc['conclusions']}" for doc in doc_texts]
|
447 |
docs_for_rerank = [self.abstract[i] for i in top_results]
|
448 |
+
|
449 |
if len(docs_for_rerank) == 0:
|
450 |
return []
|
451 |
+
|
452 |
reranked_results = self.cohere_client.rerank(
|
453 |
query=query,
|
454 |
documents=docs_for_rerank,
|
455 |
model='rerank-english-v3.0',
|
456 |
top_n=top_k
|
457 |
)
|
458 |
+
|
459 |
final_results = []
|
460 |
for result in reranked_results.results:
|
461 |
doc_id = top_results[result.index]
|
|
|
466 |
if reweight:
|
467 |
if time_result['has_temporal_aspect']:
|
468 |
final_results = self.date_filter.filter(final_results, time_score = time_result['expected_recency_weight'])
|
469 |
+
|
470 |
if self.weight_citation: self.citation_filter.filter(final_results)
|
471 |
+
|
472 |
if return_scores:
|
473 |
return {result[0]: result[2] for result in final_results}
|
474 |
|
|
|
478 |
return self.embed_batch(docs)
|
479 |
|
480 |
# ----------------------------------------------------------------
|
481 |
+
|
|
|
482 |
if 'ec' not in st.session_state:
|
483 |
+
ec = HydeCohereRetrievalSystem(weight_keywords=True)
|
484 |
st.session_state.ec = ec
|
485 |
st.toast('loaded retrieval system')
|
486 |
else:
|
487 |
ec = st.session_state.ec
|
|
|
|
|
|
|
|
|
|
|
|
|
488 |
|
489 |
+
def get_topk(query, top_k):
|
490 |
+
print('running retrieval')
|
491 |
+
rs = st.session_state.ec.retrieve(query, top_k, return_scores=True)
|
492 |
+
return rs
|
493 |
|
494 |
+
def Library(query, top_k = 7):
|
495 |
+
rs = get_topk(query, top_k = top_k)
|
496 |
+
op_docs = ''
|
497 |
+
for paperno, i in enumerate(rs):
|
498 |
+
op_docs = op_docs + 'Paper %.0f:' %(paperno+1) +' (published in '+st.session_state.arxiv_corpus['bibcode'][i][0:4] + ') ' + st.session_state.titles[i] + '\n' + st.session_state.abstracts[i] + '\n\n'
|
499 |
+
|
500 |
+
return op_docs
|
501 |
+
|
502 |
+
def Library2(query, top_k = 7):
|
503 |
+
rs = get_topk(query, top_k = top_k)
|
504 |
+
absts, fnames = [], []
|
505 |
+
for paperno, i in enumerate(rs):
|
506 |
+
absts.append(st.session_state.abstracts[i])
|
507 |
+
fnames.append(st.session_state.arxiv_corpus['bibcode'][i])
|
508 |
+
return absts, fnames, rs
|
509 |
+
|
510 |
+
def get_paper_df(ids):
|
511 |
+
|
512 |
+
papers, scores, yrs, links, cites, kws = [], [], [], [], [], []
|
513 |
for i in ids:
|
514 |
papers.append(st.session_state.titles[i])
|
515 |
scores.append(ids[i])
|
516 |
links.append('https://ui.adsabs.harvard.edu/abs/'+st.session_state.arxiv_corpus['bibcode'][i]+'/abstract')
|
517 |
+
yrs.append(st.session_state.arxiv_corpus['bibcode'][i][0:4])
|
518 |
+
cites.append(st.session_state.arxiv_corpus['cites'][i])
|
519 |
+
kws.append(st.session_state.arxiv_corpus['ads_keywords'][i])
|
520 |
+
|
521 |
return pd.DataFrame({
|
522 |
'Title': papers,
|
523 |
'Relevance': scores,
|
524 |
+
'Year': yrs,
|
525 |
+
'ADS Link': links,
|
526 |
+
'Citations': cites,
|
527 |
+
'Keywords': kws,
|
528 |
})
|
529 |
|
530 |
|
531 |
+
# def find_outliers(inp_simids, arxiv_cutoff_distance = 0.8):
|
532 |
+
#
|
533 |
+
# inp_simids = np.array(inp_simids)
|
534 |
+
#
|
535 |
+
# # Calculate the centroid for each point, excluding itself
|
536 |
+
# orange_black_points = st.session_state.embed[inp_simids]
|
537 |
+
#
|
538 |
+
# topk_dists = []
|
539 |
+
# for i, point in enumerate(orange_black_points):
|
540 |
+
# # Exclude the current point
|
541 |
+
# other_points = np.delete(orange_black_points, i, axis=0)
|
542 |
+
# # Calculate centroid of other points
|
543 |
+
# centroid = np.mean(other_points, axis=0)
|
544 |
+
# # Calculate distance from the point to this centroid
|
545 |
+
# dist = np.sqrt(np.sum((point - centroid)**2))
|
546 |
+
# topk_dists.append(dist)
|
547 |
+
#
|
548 |
+
# topk_dists = np.array(topk_dists)
|
549 |
+
#
|
550 |
+
# # Separate distances for orange and black points
|
551 |
+
# orange_distances = topk_dists[:len(inp_simids)]
|
552 |
+
# black_distances = topk_dists[len(inp_simids):]
|
553 |
+
#
|
554 |
+
# # Calculate the median of distances
|
555 |
+
# orange_black_distances = topk_dists
|
556 |
+
# median_topk_distance = np.median(orange_black_distances)
|
557 |
+
#
|
558 |
+
# # def get_sims_and_dists(inp_data):
|
559 |
+
#
|
560 |
+
# # all_sims, all_dists = [], []
|
561 |
+
#
|
562 |
+
# # np.random.seed(12)
|
563 |
+
# # rand_indices = np.random.choice(inp_data.shape[0], size=return_n, replace=False)
|
564 |
+
#
|
565 |
+
# # for j in tqdm(range(len(rand_indices))):
|
566 |
+
#
|
567 |
+
# # i = rand_indices[j]
|
568 |
+
# # inferred_vector = inp_data[i,:]
|
569 |
+
# # sims, dists = find_closest_dists(i, inp_data, return_n + 1)
|
570 |
+
# # all_sims.append(sims[1:])
|
571 |
+
# # all_dists.append(dists[1:])
|
572 |
+
#
|
573 |
+
# # return np.array(all_sims), np.array(all_dists)
|
574 |
+
#
|
575 |
+
# # # Identify papers with distances greater than the 95th percentile
|
576 |
+
# # _, all_dists = get_sims_and_dists(arxiv_ada_embeddings)
|
577 |
+
# # arxiv_cutoff_distance = find_cutoff_dist(all_dists)
|
578 |
+
# # hardcoding for now
|
579 |
+
# outlier_indices = inp_simids[np.where(orange_black_distances > arxiv_cutoff_distance)[0]]
|
580 |
+
# # outlier_titles = [titles[i] for i in outlier_indices]
|
581 |
+
#
|
582 |
+
# return outlier_indices #, outlier_titles
|
583 |
|
584 |
def create_embedding_plot(rs):
|
585 |
+
"""
|
586 |
+
function to create embedding plot
|
587 |
+
"""
|
588 |
|
589 |
pltsource = ColumnDataSource(data=dict(
|
590 |
x=st.session_state.arxiv_corpus['umap_x'],
|
|
|
592 |
title=st.session_state.titles,
|
593 |
link=st.session_state.arxiv_corpus['bibcode'],
|
594 |
))
|
595 |
+
|
596 |
rsflag = np.zeros((len(st.session_state.ids),))
|
597 |
rsflag[np.array([k for k in rs])] = 1
|
598 |
+
|
599 |
+
# outflag = np.zeros((len(st.session_state.ids),))
|
600 |
+
# outflag[np.array([k for k in find_outliers(rs)])] = 1
|
601 |
pltsource.data['colors'] = rsflag * 0.8 + 0.1
|
602 |
+
# pltsource.data['colors'][outflag] = 0.5
|
603 |
pltsource.data['sizes'] = (rsflag + 1)**5 / 100
|
604 |
|
605 |
TOOLTIPS = """
|
|
|
610 |
@link <br> <br>
|
611 |
</div>
|
612 |
"""
|
613 |
+
|
614 |
mapper = linear_cmap(field_name="colors", palette=Spectral5, low=0., high=1.)
|
615 |
|
616 |
p = figure(width=700, height=900, tooltips=TOOLTIPS, x_range=(0, 20), y_range=(-4.2,18),
|
617 |
title="UMAP projection of embeddings for the astro-ph corpus")
|
618 |
+
|
619 |
p.axis.visible=False
|
620 |
p.grid.visible=False
|
621 |
p.outline_line_alpha = 0.
|
622 |
+
|
623 |
p.circle('x', 'y', radius='sizes', source=pltsource, alpha=0.3, fill_color=mapper, fill_alpha='colors', line_color="lightgrey",line_alpha=0.1)
|
624 |
+
|
625 |
return p
|
626 |
|
627 |
+
def extract_keywords(question, ec):
|
|
|
628 |
# Simulated keyword extraction (replace with actual logic)
|
629 |
return ['keyword1', 'keyword2', 'keyword3']
|
630 |
|
|
|
633 |
# Simulated consensus estimation (replace with actual calculation)
|
634 |
return 0.75
|
635 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
636 |
|
637 |
+
def run_agent_qa(query, top_k):
|
638 |
+
|
639 |
+
# define tools
|
640 |
+
search = DuckDuckGoSearchAPIWrapper()
|
641 |
+
tools = [
|
642 |
+
Tool(
|
643 |
+
name="Library",
|
644 |
+
func=Library,
|
645 |
+
description="A source of information pertinent to your question. Do not answer a question without consulting this!"
|
646 |
+
),
|
647 |
+
Tool(
|
648 |
+
name="Search",
|
649 |
+
func=search.run,
|
650 |
+
description="useful for when you need to look up knowledge about common topics or current events",
|
651 |
+
)
|
652 |
+
]
|
653 |
|
654 |
+
if 'tools' not in st.session_state:
|
655 |
+
st.session_state.tools = tools
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
656 |
|
657 |
+
# define prompt
|
658 |
|
659 |
+
# for another question type:
|
660 |
+
# First, find the quotes from the document that are most relevant to answering the question, and then print them in numbered order.
|
661 |
+
# Quotes should be relatively short. If there are no relevant quotes, write “No relevant quotes” instead.
|
|
|
662 |
|
|
|
663 |
|
664 |
+
template = """You are an expert astronomer and cosmologist.
|
665 |
+
Answer the following question as best you can using information from the library, but speaking in a concise and factual manner.
|
666 |
+
If you can not come up with an answer, say you do not know.
|
667 |
+
Try to break the question down into smaller steps and solve it in a logical manner.
|
668 |
|
669 |
+
You have access to the following tools:
|
670 |
|
671 |
+
{tools}
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
672 |
|
673 |
+
Use the following format:
|
674 |
|
675 |
+
Question: the input question you must answer
|
676 |
+
Thought: you should always think about what to do
|
677 |
+
Action: the action to take, should be one of [{tool_names}]
|
678 |
+
Action Input: the input to the action
|
679 |
+
Observation: the result of the action
|
680 |
+
... (this Thought/Action/Action Input/Observation can repeat N times)
|
681 |
+
Thought: I now know the final answer
|
682 |
+
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
|
683 |
|
684 |
+
Begin! Remember to speak in a pedagogical and factual manner."
|
685 |
|
686 |
+
Question: {input}
|
687 |
+
Thought:{agent_scratchpad}"""
|
688 |
|
689 |
+
prompt = hub.pull("hwchase17/react")
|
690 |
+
prompt.template=template
|
691 |
+
|
692 |
+
# path to write intermediate trace to
|
693 |
+
|
694 |
+
file_path = "agent_trace.txt"
|
695 |
+
try:
|
696 |
+
os.remove(file_path)
|
697 |
+
except:
|
698 |
+
pass
|
699 |
+
file_handler = FileCallbackHandler(file_path)
|
700 |
+
callback_manager=CallbackManager([file_handler])
|
701 |
+
|
702 |
+
# define and execute agent
|
703 |
+
|
704 |
+
tool_names = [tool.name for tool in st.session_state.tools]
|
705 |
+
if 'agent' not in st.session_state:
|
706 |
+
# agent = ZeroShotAgent(llm_chain=llm_chain, allowed_tools=tool_names)
|
707 |
+
agent = create_react_agent(llm=gen_llm, tools=tools, prompt=prompt)
|
708 |
+
st.session_state.agent = agent
|
709 |
+
|
710 |
+
if 'agent_executor' not in st.session_state:
|
711 |
+
agent_executor = AgentExecutor(agent=st.session_state.agent, tools=st.session_state.tools, verbose=True, handle_parsing_errors=True, callbacks=CallbackManager([file_handler]))
|
712 |
+
st.session_state.agent_executor = agent_executor
|
713 |
+
|
714 |
+
answer = st.session_state.agent_executor.invoke({"input": query,})
|
715 |
+
return answer
|
716 |
+
|
717 |
+
def make_rag_qa_answer(query, top_k = 10):
|
718 |
+
|
719 |
+
absts, fhdrs, rs = Library2(query, top_k = top_k)
|
720 |
+
|
721 |
+
temp_abst = ''
|
722 |
+
loaders = []
|
723 |
+
for i in range(len(absts)):
|
724 |
+
temp_abst = absts[i]
|
725 |
+
|
726 |
+
try:
|
727 |
+
text_file = open("absts/"+fhdrs[i]+".txt", "w")
|
728 |
+
except:
|
729 |
+
os.mkdir('absts')
|
730 |
+
text_file = open("absts/"+fhdrs[i]+".txt", "w")
|
731 |
+
n = text_file.write(temp_abst)
|
732 |
+
text_file.close()
|
733 |
+
loader = TextLoader("absts/"+fhdrs[i]+".txt")
|
734 |
+
loaders.append(loader)
|
735 |
+
|
736 |
+
text_splitter = RecursiveCharacterTextSplitter(chunk_size=150, chunk_overlap=50, add_start_index=True)
|
737 |
+
|
738 |
+
splits = text_splitter.split_documents([loader.load()[0] for loader in loaders])
|
739 |
+
vectorstore = Chroma.from_documents(documents=splits, embedding=embeddings, collection_name='retdoc4')
|
740 |
+
# retriever = vectorstore.as_retriever(search_type="similarity", search_kwargs={"k": 6, "fetch_k": len(splits)})
|
741 |
+
retriever = vectorstore.as_retriever(search_type="similarity", search_kwargs={"k": 6})
|
742 |
+
|
743 |
+
for i in range(len(absts)):
|
744 |
+
os.remove("absts/"+fhdrs[i]+".txt")
|
745 |
+
|
746 |
+
template = """You are an expert astronomer and cosmologist.
|
747 |
+
Answer the following question as best you can using information from the library, but speaking in a concise and factual manner.
|
748 |
+
If you can not come up with an answer, say you do not know.
|
749 |
+
Try to break the question down into smaller steps and solve it in a logical manner.
|
750 |
+
|
751 |
+
Provide information about how you arrived at the answer, and any nuances or uncertainties the reader should be aware of.
|
752 |
+
|
753 |
+
Begin! Remember to speak in a pedagogical and factual manner."
|
754 |
+
|
755 |
+
Relevant documents:{context}
|
756 |
|
757 |
+
Question: {question}
|
758 |
+
Answer:"""
|
759 |
+
prompt = PromptTemplate.from_template(template)
|
|
|
|
|
760 |
|
761 |
+
def format_docs(docs):
|
762 |
+
return "\n\n".join(doc.page_content for doc in docs)
|
763 |
+
|
764 |
+
|
765 |
+
rag_chain_from_docs = (
|
766 |
+
RunnablePassthrough.assign(context=(lambda x: format_docs(x["context"])))
|
767 |
+
| prompt
|
768 |
+
| gen_llm
|
769 |
+
| StrOutputParser()
|
770 |
+
)
|
771 |
+
|
772 |
+
rag_chain_with_source = RunnableParallel(
|
773 |
+
{"context": retriever, "question": RunnablePassthrough()}
|
774 |
+
).assign(answer=rag_chain_from_docs)
|
775 |
+
|
776 |
+
rag_answer = rag_chain_with_source.invoke(query, )
|
777 |
+
|
778 |
+
vectorstore.delete_collection()
|
779 |
+
return rag_answer, rs
|
780 |
+
|
781 |
+
def guess_question_type(query: str):
|
782 |
+
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:
|
783 |
+
|
784 |
+
1. Summarization
|
785 |
+
2. Single-paper factual
|
786 |
+
3. Multi-paper factual
|
787 |
+
4. Named entity recognition
|
788 |
+
5. Jargon-specific questions / overloaded words
|
789 |
+
6. Time-sensitive
|
790 |
+
7. Consensus evaluation
|
791 |
+
8. What-ifs and counterfactuals
|
792 |
+
9. Compositional
|
793 |
+
|
794 |
+
Analyze the query carefully, considering its content, structure, and implications. Then, determine which of the above categories best fits the query.
|
795 |
+
|
796 |
+
In your analysis, consider the following:
|
797 |
+
- Does the query ask for a well-known datapoint or mechanism?
|
798 |
+
- Can it be answered by a single paper or does it require multiple sources?
|
799 |
+
- Does it involve proper nouns or specific scientific terms?
|
800 |
+
- Is it time-dependent or likely to change in the near future?
|
801 |
+
- Does it require evaluating consensus across multiple sources?
|
802 |
+
- Is it a hypothetical or counterfactual question?
|
803 |
+
- Does it need to be broken down into sub-queries (i.e. compositional)?
|
804 |
+
|
805 |
+
After your analysis, categorize the query into one of the nine categories listed above.
|
806 |
+
|
807 |
+
Provide a brief explanation for your categorization, highlighting the key aspects of the query that led to your decision.
|
808 |
+
|
809 |
+
Present your final answer in the following format:
|
810 |
+
|
811 |
+
<categorization>
|
812 |
+
Category: [Selected category]
|
813 |
+
Explanation: [Your explanation for the categorization]
|
814 |
+
</categorization>"""
|
815 |
+
# st.write('invoking hyde generation')
|
816 |
+
|
817 |
+
# message = self.generation_client.messages.create(
|
818 |
+
# model = self.generation_model,
|
819 |
+
# max_tokens = self.max_doclen,
|
820 |
+
# temperature = self.temperature,
|
821 |
+
# system = prompt,
|
822 |
+
# messages=[{ "role": "user",
|
823 |
+
# "content": [{"type": "text", "text": query,}] }]
|
824 |
+
# )
|
825 |
+
# return message.content[0].text
|
826 |
+
|
827 |
+
messages = [("system",categorization_prompt,),("human", query),]
|
828 |
+
return st.session_state.ec.generation_client.invoke(messages).content
|
829 |
+
|
830 |
+
|
831 |
+
class OverallConsensusEvaluation(BaseModel):
|
832 |
+
consensus: Literal["Strong Agreement", "Moderate Agreement", "Weak Agreement", "No Clear Consensus", "Weak Disagreement", "Moderate Disagreement", "Strong Disagreement"] = Field(
|
833 |
+
...,
|
834 |
+
description="The overall level of consensus between the query and the abstracts"
|
835 |
+
)
|
836 |
+
explanation: str = Field(
|
837 |
+
...,
|
838 |
+
description="A detailed explanation of the consensus evaluation"
|
839 |
+
)
|
840 |
+
relevance_score: float = Field(
|
841 |
+
...,
|
842 |
+
description="A score from 0 to 1 indicating how relevant the abstracts are to the query overall",
|
843 |
+
ge=0,
|
844 |
+
le=1
|
845 |
+
)
|
846 |
+
|
847 |
+
def evaluate_overall_consensus(query: str, abstracts: List[str]) -> OverallConsensusEvaluation:
|
848 |
+
"""
|
849 |
+
Evaluates the overall consensus of the abstracts in relation to the query in a single LLM call.
|
850 |
+
"""
|
851 |
+
prompt = f"""
|
852 |
+
Query: {query}
|
853 |
+
|
854 |
+
You will be provided with {len(abstracts)} scientific abstracts. Your task is to:
|
855 |
+
1. Evaluate the overall consensus between the query and the abstracts.
|
856 |
+
2. Provide a detailed explanation of your consensus evaluation.
|
857 |
+
3. Assign an overall relevance score from 0 to 1, where 0 means completely irrelevant and 1 means highly relevant.
|
858 |
+
|
859 |
+
For the consensus evaluation, use one of the following levels:
|
860 |
+
Strong Agreement, Moderate Agreement, Weak Agreement, No Clear Consensus, Weak Disagreement, Moderate Disagreement, Strong Disagreement
|
861 |
+
|
862 |
+
Here are the abstracts:
|
863 |
+
|
864 |
+
{' '.join([f"Abstract {i+1}: {abstract}" for i, abstract in enumerate(abstracts)])}
|
865 |
+
|
866 |
+
Provide your evaluation in a structured format.
|
867 |
+
"""
|
868 |
+
|
869 |
+
response = consensus_client.chat.completions.create(
|
870 |
+
model="gpt-4",
|
871 |
+
response_model=OverallConsensusEvaluation,
|
872 |
+
messages=[
|
873 |
+
{"role": "system", "content": """You are an assistant with expertise in astrophysics for question-answering tasks.
|
874 |
+
Evaluate the overall consensus of the retrieved scientific abstracts in relation to a given query.
|
875 |
+
If you don't know the answer, just say that you don't know.
|
876 |
+
Use six sentences maximum and keep the answer concise."""},
|
877 |
+
{"role": "user", "content": prompt}
|
878 |
+
],
|
879 |
+
temperature=0
|
880 |
+
)
|
881 |
+
|
882 |
+
return response
|
883 |
|
884 |
|
885 |
# Streamlit app
|
886 |
def main():
|
887 |
+
|
888 |
# st.title("Question Answering App")
|
889 |
+
|
890 |
+
|
891 |
# Sidebar (Inputs)
|
892 |
st.sidebar.header("Fine-tune the search")
|
893 |
top_k = st.sidebar.slider("Number of papers to retrieve:", 3, 30, 10)
|
894 |
extra_keywords = st.sidebar.text_input("Enter extra keywords (comma-separated):")
|
895 |
+
|
896 |
st.sidebar.subheader("Toggles")
|
897 |
+
toggle_a = st.sidebar.toggle("Weight by keywords", value = False)
|
898 |
+
toggle_b = st.sidebar.toggle("Weight by date", value = False)
|
899 |
+
toggle_c = st.sidebar.toggle("Weight by citations", value = False)
|
900 |
+
|
901 |
+
method = st.sidebar.radio("Retrieval method:", ["Semantic search", "Semantic search + HyDE", "Semantic search + HyDE + CoHERE"], index=2)
|
902 |
+
if (method == "Semantic search"):
|
903 |
+
with st.spinner('set retrieval method to'+ method):
|
904 |
+
st.session_state.ec = EmbeddingRetrievalSystem(weight_keywords=True)
|
905 |
+
elif (method == "Semantic search + HyDE"):
|
906 |
+
with st.spinner('set retrieval method to'+ method):
|
907 |
+
st.session_state.ec = HydeRetrievalSystem(weight_keywords=True)
|
908 |
+
elif (method == "Semantic search + HyDE + CoHERE"):
|
909 |
+
with st.spinner('set retrieval method to'+ method):
|
910 |
+
st.session_state.ec = HydeCohereRetrievalSystem(weight_keywords=True)
|
911 |
+
|
912 |
+
method2 = st.sidebar.radio("Generation complexity:", ["Basic RAG","ReAct Agent"])
|
913 |
+
if method2 == "Basic RAG":
|
914 |
+
st.session_state.gen_method = 'rag'
|
915 |
+
elif method2 == "ReAct Agent":
|
916 |
+
st.session_state.gen_method = 'agent'
|
917 |
+
|
918 |
|
919 |
+
question_type = st.sidebar.selectbox("Select question type:", ["Single paper", "Multi-paper", "Summary"])
|
920 |
store_output = st.sidebar.button("Save output")
|
921 |
|
922 |
# Main page (Outputs)
|
923 |
+
# st.markdown("""
|
924 |
+
# <style>
|
925 |
+
# .stTextInput > div > div { font-size: 50px; }
|
926 |
+
# </style>
|
927 |
+
# """, unsafe_allow_html=True)
|
928 |
+
|
929 |
+
# st.markdown(
|
930 |
+
# """
|
931 |
+
# <style>
|
932 |
+
# textarea {
|
933 |
+
# font-size: 3rem !important;
|
934 |
+
# font-weight: bold;
|
935 |
+
# font-family: "Times New Roman", Times, serif;
|
936 |
+
# }
|
937 |
+
# input {
|
938 |
+
# font-size: 3rem !important;
|
939 |
+
# font-weight: bold;
|
940 |
+
# font-family: "Times New Roman", Times, serif;
|
941 |
+
# }
|
942 |
+
# </style>
|
943 |
+
# """,
|
944 |
+
# unsafe_allow_html=True,
|
945 |
+
# )
|
946 |
+
# query = st.text_area("Ask me anything:", height=30)
|
947 |
+
|
948 |
query = st.text_input("Ask me anything:")
|
949 |
submit_button = st.button("Submit")
|
950 |
+
|
951 |
if submit_button:
|
952 |
+
|
953 |
+
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...']
|
954 |
+
|
955 |
+
with st.spinner(search_text_list[np.random.choice(len(search_text_list))]):
|
956 |
+
|
957 |
+
# Process inputs
|
958 |
+
keywords = [kw.strip() for kw in extra_keywords.split(',')] if extra_keywords else []
|
959 |
+
toggles = {'Keyword weighting': toggle_a, 'Time weighting': toggle_b, 'Citation weighting': toggle_c}
|
960 |
+
# Generate outputs
|
961 |
+
|
962 |
+
st.session_state.ec.query_input_keywords = keywords
|
963 |
+
st.session_state.ec.toggles = toggles
|
964 |
+
st.session_state.ec.question_type = question_type
|
965 |
+
st.session_state.ec.rag_method = method
|
966 |
+
st.session_state.ec.gen_method = method2
|
967 |
+
|
968 |
+
# Display outputs
|
969 |
+
if st.session_state.gen_method == 'agent':
|
970 |
+
answer = run_agent_qa(query, top_k)
|
971 |
+
rs = get_topk(query, top_k)
|
972 |
+
|
973 |
+
st.write(answer["output"])
|
974 |
+
|
975 |
+
file_path = "agent_trace.txt"
|
976 |
+
with open(file_path, 'r') as file:
|
977 |
+
intermediate_steps = file.read()
|
978 |
+
|
979 |
+
st.expander('Intermediate steps', expanded=False).write(intermediate_steps)
|
980 |
+
|
981 |
+
elif st.session_state.gen_method == 'rag':
|
982 |
+
answer, rs = make_rag_qa_answer(query, top_k)
|
983 |
+
st.write(answer['answer'])
|
984 |
+
|
985 |
+
papers_df = get_paper_df(rs)
|
986 |
+
embedding_plot = create_embedding_plot(rs)
|
987 |
+
triggered_keywords = st.session_state.ec.query_kws
|
988 |
+
st.write('**Triggered keywords:** `'+ "`, `".join(triggered_keywords)+'`')
|
989 |
+
# consensus = estimate_consensus()
|
990 |
+
|
991 |
+
|
992 |
with st.expander("Relevant papers", expanded=True):
|
993 |
# st.dataframe(papers_df, hide_index=True)
|
994 |
st.data_editor(papers_df,
|
995 |
+
column_config = {'ADS Link':st.column_config.LinkColumn(display_text= 'https://ui.adsabs.harvard.edu/abs/(.*?)/abstract')}
|
996 |
)
|
997 |
|
998 |
with st.expander("Embedding map", expanded=False):
|
999 |
st.bokeh_chart(embedding_plot)
|
1000 |
+
|
1001 |
col1, col2 = st.columns(2)
|
1002 |
+
|
1003 |
with col1:
|
1004 |
+
|
1005 |
+
st.subheader("Question type suggestion")
|
1006 |
+
question_type_gen = guess_question_type(query)
|
1007 |
+
if '<categorization>' in question_type_gen:
|
1008 |
+
question_type_gen = question_type_gen.split('<categorization>')[1]
|
1009 |
+
if '</categorization>' in question_type_gen:
|
1010 |
+
question_type_gen = question_type_gen.split('</categorization>')[0]
|
1011 |
+
question_type_gen = question_type_gen.replace('\n',' \n')
|
1012 |
+
st.markdown(question_type_gen)
|
1013 |
+
|
1014 |
with col2:
|
1015 |
+
|
1016 |
+
# st.subheader("Triggered Keywords")
|
1017 |
+
# st.write(", ".join(triggered_keywords))
|
1018 |
+
|
1019 |
+
consensus_answer = evaluate_overall_consensus(query, [st.session_state.abstracts[i] for i in rs])
|
1020 |
+
st.subheader("Consensus: "+consensus_answer.consensus)
|
1021 |
+
st.markdown(consensus_answer.explanation)
|
1022 |
+
st.markdown('Relevance of retrieved papers to answer: %.1f' %consensus_answer.relevance_score)
|
1023 |
+
|
1024 |
+
# st.write(f"{consensus:.2%}")
|
1025 |
+
|
1026 |
else:
|
1027 |
+
st.info("Use the sidebar to tweak the search parameters to get better results.")
|
1028 |
+
|
1029 |
if store_output:
|
1030 |
st.toast("Output stored successfully!")
|
1031 |
|
1032 |
if __name__ == "__main__":
|
1033 |
+
main()
|
data/data-00000-of-00001.arrow
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:2534048757c630a5a9addf362d3077da0427e55ae1cae0c93dd213363ddfbcc7
|
3 |
+
size 498031096
|
data/dataset_info.json
ADDED
@@ -0,0 +1,134 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
{
|
2 |
+
"builder_name": "parquet",
|
3 |
+
"citation": "",
|
4 |
+
"config_name": "default",
|
5 |
+
"dataset_name": "pathfinder_arxiv_data_galaxy",
|
6 |
+
"dataset_size": 505886100,
|
7 |
+
"description": "",
|
8 |
+
"download_checksums": {
|
9 |
+
"hf://datasets/kiyer/pathfinder_arxiv_data_galaxy@29754b03f3cd82e4051ece1cf96605f8756bc197/data/train-00000-of-00001.parquet": {
|
10 |
+
"num_bytes": 379674094,
|
11 |
+
"checksum": null
|
12 |
+
}
|
13 |
+
},
|
14 |
+
"download_size": 379674094,
|
15 |
+
"features": {
|
16 |
+
"ads_id": {
|
17 |
+
"dtype": "string",
|
18 |
+
"_type": "Value"
|
19 |
+
},
|
20 |
+
"arxiv_id": {
|
21 |
+
"dtype": "string",
|
22 |
+
"_type": "Value"
|
23 |
+
},
|
24 |
+
"title": {
|
25 |
+
"dtype": "string",
|
26 |
+
"_type": "Value"
|
27 |
+
},
|
28 |
+
"abstract": {
|
29 |
+
"dtype": "string",
|
30 |
+
"_type": "Value"
|
31 |
+
},
|
32 |
+
"embed": {
|
33 |
+
"feature": {
|
34 |
+
"dtype": "float32",
|
35 |
+
"_type": "Value"
|
36 |
+
},
|
37 |
+
"_type": "Sequence"
|
38 |
+
},
|
39 |
+
"umap_x": {
|
40 |
+
"dtype": "float32",
|
41 |
+
"_type": "Value"
|
42 |
+
},
|
43 |
+
"umap_y": {
|
44 |
+
"dtype": "float32",
|
45 |
+
"_type": "Value"
|
46 |
+
},
|
47 |
+
"date": {
|
48 |
+
"dtype": "date32",
|
49 |
+
"_type": "Value"
|
50 |
+
},
|
51 |
+
"cites": {
|
52 |
+
"dtype": "int64",
|
53 |
+
"_type": "Value"
|
54 |
+
},
|
55 |
+
"bibcode": {
|
56 |
+
"dtype": "string",
|
57 |
+
"_type": "Value"
|
58 |
+
},
|
59 |
+
"keywords": {
|
60 |
+
"feature": {
|
61 |
+
"dtype": "string",
|
62 |
+
"_type": "Value"
|
63 |
+
},
|
64 |
+
"_type": "Sequence"
|
65 |
+
},
|
66 |
+
"ads_keywords": {
|
67 |
+
"feature": {
|
68 |
+
"dtype": "string",
|
69 |
+
"_type": "Value"
|
70 |
+
},
|
71 |
+
"_type": "Sequence"
|
72 |
+
},
|
73 |
+
"read_count": {
|
74 |
+
"dtype": "int64",
|
75 |
+
"_type": "Value"
|
76 |
+
},
|
77 |
+
"doi": {
|
78 |
+
"feature": {
|
79 |
+
"dtype": "string",
|
80 |
+
"_type": "Value"
|
81 |
+
},
|
82 |
+
"_type": "Sequence"
|
83 |
+
},
|
84 |
+
"authors": {
|
85 |
+
"feature": {
|
86 |
+
"dtype": "string",
|
87 |
+
"_type": "Value"
|
88 |
+
},
|
89 |
+
"_type": "Sequence"
|
90 |
+
},
|
91 |
+
"aff": {
|
92 |
+
"feature": {
|
93 |
+
"dtype": "string",
|
94 |
+
"_type": "Value"
|
95 |
+
},
|
96 |
+
"_type": "Sequence"
|
97 |
+
},
|
98 |
+
"cite_bibcodes": {
|
99 |
+
"feature": {
|
100 |
+
"dtype": "string",
|
101 |
+
"_type": "Value"
|
102 |
+
},
|
103 |
+
"_type": "Sequence"
|
104 |
+
},
|
105 |
+
"ref_bibcodes": {
|
106 |
+
"feature": {
|
107 |
+
"dtype": "string",
|
108 |
+
"_type": "Value"
|
109 |
+
},
|
110 |
+
"_type": "Sequence"
|
111 |
+
}
|
112 |
+
},
|
113 |
+
"homepage": "",
|
114 |
+
"license": "",
|
115 |
+
"size_in_bytes": 885560194,
|
116 |
+
"splits": {
|
117 |
+
"train": {
|
118 |
+
"name": "train",
|
119 |
+
"num_bytes": 505886100,
|
120 |
+
"num_examples": 41195,
|
121 |
+
"shard_lengths": [
|
122 |
+
41000,
|
123 |
+
195
|
124 |
+
],
|
125 |
+
"dataset_name": "pathfinder_arxiv_data_galaxy"
|
126 |
+
}
|
127 |
+
},
|
128 |
+
"version": {
|
129 |
+
"version_str": "0.0.0",
|
130 |
+
"major": 0,
|
131 |
+
"minor": 0,
|
132 |
+
"patch": 0
|
133 |
+
}
|
134 |
+
}
|
data/state.json
ADDED
@@ -0,0 +1,13 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
{
|
2 |
+
"_data_files": [
|
3 |
+
{
|
4 |
+
"filename": "data-00000-of-00001.arrow"
|
5 |
+
}
|
6 |
+
],
|
7 |
+
"_fingerprint": "61bcd9aec14a17d4",
|
8 |
+
"_format_columns": null,
|
9 |
+
"_format_kwargs": {},
|
10 |
+
"_format_type": null,
|
11 |
+
"_output_all_columns": false,
|
12 |
+
"_split": "train"
|
13 |
+
}
|
requirements.txt
CHANGED
@@ -10,7 +10,8 @@ langchain_community
|
|
10 |
langchain_core
|
11 |
langchainhub
|
12 |
openai
|
13 |
-
|
|
|
14 |
feedparser
|
15 |
tiktoken
|
16 |
chromadb
|
|
|
10 |
langchain_core
|
11 |
langchainhub
|
12 |
openai
|
13 |
+
instructor
|
14 |
+
pydantic
|
15 |
feedparser
|
16 |
tiktoken
|
17 |
chromadb
|