lfoppiano commited on
Commit
8700e22
β€’
2 Parent(s): a70fbd3 7b33dad

Merge branch 'main' into pdf-render

Browse files
CHANGELOG.md CHANGED
@@ -4,27 +4,49 @@ All notable changes to this project will be documented in this file.
4
 
5
  The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.0.0/).
6
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
7
 
8
  ## [0.2.0] – 2023-10-31
9
 
10
  ### Added
 
11
  + Selection of chunk size on which embeddings are created upon
12
- + Mistral model to be used freely via the Huggingface free API
13
 
14
  ### Changed
15
- + Improved documentation, adding privacy statement
 
16
  + Moved settings on the sidebar
17
  + Disable NER extraction by default, and allow user to activate it
18
  + Read API KEY from the environment variables and if present, avoid asking the user
19
  + Avoid changing model after update
20
 
21
-
22
-
23
  ## [0.1.3] – 2023-10-30
24
 
25
  ### Fixed
26
 
27
- + ChromaDb accumulating information even when new papers were uploaded
28
 
29
  ## [0.1.2] – 2023-10-26
30
 
@@ -36,9 +58,8 @@ The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.0.0/).
36
 
37
  ### Fixed
38
 
39
- + Github action build
40
- + dependencies of langchain and chromadb
41
-
42
 
43
  ## [0.1.0] – 2023-10-26
44
 
@@ -54,8 +75,8 @@ The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.0.0/).
54
  + Kick off application
55
  + Support for GPT-3.5
56
  + Support for Mistral + SentenceTransformer
57
- + Streamlit application
58
- + Docker image
59
  + pypi package
60
 
61
  <!-- markdownlint-disable-file MD024 MD033 -->
 
4
 
5
  The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.0.0/).
6
 
7
+ ## [0.3.1] - 2023-11-22
8
+
9
+ ### Added
10
+
11
+ + Include biblio in embeddings by @lfoppiano in #21
12
+
13
+ ### Fixed
14
+
15
+ + Fix conversational memory by @lfoppiano in #20
16
+
17
+ ## [0.3.0] - 2023-11-18
18
+
19
+ ### Added
20
+
21
+ + add zephyr-7b by @lfoppiano in #15
22
+ + add conversational memory in #18
23
+
24
+ ## [0.2.1] - 2023-11-01
25
+
26
+ ### Fixed
27
+
28
+ + fix env variables by @lfoppiano in #9
29
 
30
  ## [0.2.0] – 2023-10-31
31
 
32
  ### Added
33
+
34
  + Selection of chunk size on which embeddings are created upon
35
+ + Mistral model to be used freely via the Huggingface free API
36
 
37
  ### Changed
38
+
39
+ + Improved documentation, adding privacy statement
40
  + Moved settings on the sidebar
41
  + Disable NER extraction by default, and allow user to activate it
42
  + Read API KEY from the environment variables and if present, avoid asking the user
43
  + Avoid changing model after update
44
 
 
 
45
  ## [0.1.3] – 2023-10-30
46
 
47
  ### Fixed
48
 
49
+ + ChromaDb accumulating information even when new papers were uploaded
50
 
51
  ## [0.1.2] – 2023-10-26
52
 
 
58
 
59
  ### Fixed
60
 
61
+ + Github action build
62
+ + dependencies of langchain and chromadb
 
63
 
64
  ## [0.1.0] – 2023-10-26
65
 
 
75
  + Kick off application
76
  + Support for GPT-3.5
77
  + Support for Mistral + SentenceTransformer
78
+ + Streamlit application
79
+ + Docker image
80
  + pypi package
81
 
82
  <!-- markdownlint-disable-file MD024 MD033 -->
README.md CHANGED
@@ -14,6 +14,8 @@ license: apache-2.0
14
 
15
  **Work in progress** :construction_worker:
16
 
 
 
17
  ## Introduction
18
 
19
  Question/Answering on scientific documents using LLMs: ChatGPT-3.5-turbo, Mistral-7b-instruct and Zephyr-7b-beta.
@@ -23,11 +25,13 @@ We target only the full-text using [Grobid](https://github.com/kermitt2/grobid)
23
 
24
  Additionally, this frontend provides the visualisation of named entities on LLM responses to extract <span stype="color:yellow">physical quantities, measurements</span> (with [grobid-quantities](https://github.com/kermitt2/grobid-quantities)) and <span stype="color:blue">materials</span> mentions (with [grobid-superconductors](https://github.com/lfoppiano/grobid-superconductors)).
25
 
26
- The conversation is backed up by a sliding window memory (top 4 more recent messages) that help refers to information previously discussed in the chat.
 
 
27
 
28
  **Demos**:
29
- - (on HuggingFace spaces): https://lfoppiano-document-qa.hf.space/
30
- - (on the Streamlit cloud): https://document-insights.streamlit.app/
31
 
32
  ## Getting started
33
 
 
14
 
15
  **Work in progress** :construction_worker:
16
 
17
+ <img src="https://github.com/lfoppiano/document-qa/assets/15426/f0a04a86-96b3-406e-8303-904b93f00015" width=300 align="right" />
18
+
19
  ## Introduction
20
 
21
  Question/Answering on scientific documents using LLMs: ChatGPT-3.5-turbo, Mistral-7b-instruct and Zephyr-7b-beta.
 
25
 
26
  Additionally, this frontend provides the visualisation of named entities on LLM responses to extract <span stype="color:yellow">physical quantities, measurements</span> (with [grobid-quantities](https://github.com/kermitt2/grobid-quantities)) and <span stype="color:blue">materials</span> mentions (with [grobid-superconductors](https://github.com/lfoppiano/grobid-superconductors)).
27
 
28
+ The conversation is kept in memory up by a buffered sliding window memory (top 4 more recent messages) and the messages are injected in the context as "previous messages".
29
+
30
+ (The image on the right was generated with https://huggingface.co/spaces/stabilityai/stable-diffusion)
31
 
32
  **Demos**:
33
+ - (stable version): https://lfoppiano-document-qa.hf.space/
34
+ - (unstable version): https://document-insights.streamlit.app/
35
 
36
  ## Getting started
37
 
document_qa/document_qa_engine.py CHANGED
@@ -3,17 +3,18 @@ import os
3
  from pathlib import Path
4
  from typing import Union, Any
5
 
 
6
  from grobid_client.grobid_client import GrobidClient
7
- from langchain.chains import create_extraction_chain
8
- from langchain.chains.question_answering import load_qa_chain
 
9
  from langchain.prompts import SystemMessagePromptTemplate, HumanMessagePromptTemplate, ChatPromptTemplate
10
  from langchain.retrievers import MultiQueryRetriever
 
11
  from langchain.text_splitter import RecursiveCharacterTextSplitter
12
  from langchain.vectorstores import Chroma
13
  from tqdm import tqdm
14
 
15
- from document_qa.grobid_processors import GrobidProcessor
16
-
17
 
18
  class DocumentQAEngine:
19
  llm = None
@@ -23,15 +24,24 @@ class DocumentQAEngine:
23
  embeddings_map_from_md5 = {}
24
  embeddings_map_to_md5 = {}
25
 
 
 
 
 
 
 
 
26
  def __init__(self,
27
  llm,
28
  embedding_function,
29
  qa_chain_type="stuff",
30
  embeddings_root_path=None,
31
  grobid_url=None,
 
32
  ):
33
  self.embedding_function = embedding_function
34
  self.llm = llm
 
35
  self.chain = load_qa_chain(llm, chain_type=qa_chain_type)
36
 
37
  if embeddings_root_path is not None:
@@ -87,14 +97,14 @@ class DocumentQAEngine:
87
  return self.embeddings_map_from_md5[md5]
88
 
89
  def query_document(self, query: str, doc_id, output_parser=None, context_size=4, extraction_schema=None,
90
- verbose=False, memory=None) -> (
91
  Any, str):
92
  # self.load_embeddings(self.embeddings_root_path)
93
 
94
  if verbose:
95
  print(query)
96
 
97
- response = self._run_query(doc_id, query, context_size=context_size, memory=memory)
98
  response = response['output_text'] if 'output_text' in response else response
99
 
100
  if verbose:
@@ -144,21 +154,25 @@ class DocumentQAEngine:
144
 
145
  return parsed_output
146
 
147
- def _run_query(self, doc_id, query, memory=None, context_size=4):
148
  relevant_documents = self._get_context(doc_id, query, context_size)
149
- if memory:
150
- return self.chain.run(input_documents=relevant_documents,
151
  question=query)
152
- else:
153
- return self.chain.run(input_documents=relevant_documents,
154
- question=query,
155
- memory=memory)
156
- # return self.chain({"input_documents": relevant_documents, "question": prompt_chat_template}, return_only_outputs=True)
157
 
158
  def _get_context(self, doc_id, query, context_size=4):
159
  db = self.embeddings_dict[doc_id]
160
  retriever = db.as_retriever(search_kwargs={"k": context_size})
161
  relevant_documents = retriever.get_relevant_documents(query)
 
 
 
 
 
 
162
  return relevant_documents
163
 
164
  def get_all_context_by_document(self, doc_id):
@@ -173,8 +187,10 @@ class DocumentQAEngine:
173
  relevant_documents = multi_query_retriever.get_relevant_documents(query)
174
  return relevant_documents
175
 
176
- def get_text_from_document(self, pdf_file_path, chunk_size=-1, perc_overlap=0.1, verbose=False):
177
- """Extract text from documents using Grobid, if chunk_size is < 0 it keep each paragraph separately"""
 
 
178
  if verbose:
179
  print("File", pdf_file_path)
180
  filename = Path(pdf_file_path).stem
@@ -189,6 +205,7 @@ class DocumentQAEngine:
189
  texts = []
190
  metadatas = []
191
  ids = []
 
192
  if chunk_size < 0:
193
  for passage in structure['passages']:
194
  biblio_copy = copy.copy(biblio)
@@ -212,28 +229,49 @@ class DocumentQAEngine:
212
  metadatas = [biblio for _ in range(len(texts))]
213
  ids = [id for id, t in enumerate(texts)]
214
 
 
 
 
 
 
 
 
 
 
 
215
  return texts, metadatas, ids
216
 
217
- def create_memory_embeddings(self, pdf_path, doc_id=None, chunk_size=500, perc_overlap=0.1):
218
- texts, metadata, ids = self.get_text_from_document(pdf_path, chunk_size=chunk_size, perc_overlap=perc_overlap)
 
 
 
 
 
219
  if doc_id:
220
  hash = doc_id
221
  else:
222
  hash = metadata[0]['hash']
223
 
224
  if hash not in self.embeddings_dict.keys():
225
- self.embeddings_dict[hash] = Chroma.from_texts(texts, embedding=self.embedding_function, metadatas=metadata,
 
 
226
  collection_name=hash)
227
  else:
228
- self.embeddings_dict[hash].delete(ids=self.embeddings_dict[hash].get()['ids'])
229
- self.embeddings_dict[hash] = Chroma.from_texts(texts, embedding=self.embedding_function, metadatas=metadata,
 
 
 
 
230
  collection_name=hash)
231
 
232
  self.embeddings_root_path = None
233
 
234
  return hash
235
 
236
- def create_embeddings(self, pdfs_dir_path: Path, chunk_size=500, perc_overlap=0.1):
237
  input_files = []
238
  for root, dirs, files in os.walk(pdfs_dir_path, followlinks=False):
239
  for file_ in files:
@@ -250,9 +288,12 @@ class DocumentQAEngine:
250
  if os.path.exists(data_path):
251
  print(data_path, "exists. Skipping it ")
252
  continue
253
-
254
- texts, metadata, ids = self.get_text_from_document(input_file, chunk_size=chunk_size,
255
- perc_overlap=perc_overlap)
 
 
 
256
  filename = metadata[0]['filename']
257
 
258
  vector_db_document = Chroma.from_texts(texts,
 
3
  from pathlib import Path
4
  from typing import Union, Any
5
 
6
+ from document_qa.grobid_processors import GrobidProcessor
7
  from grobid_client.grobid_client import GrobidClient
8
+ from langchain.chains import create_extraction_chain, ConversationChain, ConversationalRetrievalChain
9
+ from langchain.chains.question_answering import load_qa_chain, stuff_prompt, refine_prompts, map_reduce_prompt, \
10
+ map_rerank_prompt
11
  from langchain.prompts import SystemMessagePromptTemplate, HumanMessagePromptTemplate, ChatPromptTemplate
12
  from langchain.retrievers import MultiQueryRetriever
13
+ from langchain.schema import Document
14
  from langchain.text_splitter import RecursiveCharacterTextSplitter
15
  from langchain.vectorstores import Chroma
16
  from tqdm import tqdm
17
 
 
 
18
 
19
  class DocumentQAEngine:
20
  llm = None
 
24
  embeddings_map_from_md5 = {}
25
  embeddings_map_to_md5 = {}
26
 
27
+ default_prompts = {
28
+ 'stuff': stuff_prompt,
29
+ 'refine': refine_prompts,
30
+ "map_reduce": map_reduce_prompt,
31
+ "map_rerank": map_rerank_prompt
32
+ }
33
+
34
  def __init__(self,
35
  llm,
36
  embedding_function,
37
  qa_chain_type="stuff",
38
  embeddings_root_path=None,
39
  grobid_url=None,
40
+ memory=None
41
  ):
42
  self.embedding_function = embedding_function
43
  self.llm = llm
44
+ self.memory = memory
45
  self.chain = load_qa_chain(llm, chain_type=qa_chain_type)
46
 
47
  if embeddings_root_path is not None:
 
97
  return self.embeddings_map_from_md5[md5]
98
 
99
  def query_document(self, query: str, doc_id, output_parser=None, context_size=4, extraction_schema=None,
100
+ verbose=False) -> (
101
  Any, str):
102
  # self.load_embeddings(self.embeddings_root_path)
103
 
104
  if verbose:
105
  print(query)
106
 
107
+ response = self._run_query(doc_id, query, context_size=context_size)
108
  response = response['output_text'] if 'output_text' in response else response
109
 
110
  if verbose:
 
154
 
155
  return parsed_output
156
 
157
+ def _run_query(self, doc_id, query, context_size=4):
158
  relevant_documents = self._get_context(doc_id, query, context_size)
159
+ response = self.chain.run(input_documents=relevant_documents,
 
160
  question=query)
161
+
162
+ if self.memory:
163
+ self.memory.save_context({"input": query}, {"output": response})
164
+ return response
 
165
 
166
  def _get_context(self, doc_id, query, context_size=4):
167
  db = self.embeddings_dict[doc_id]
168
  retriever = db.as_retriever(search_kwargs={"k": context_size})
169
  relevant_documents = retriever.get_relevant_documents(query)
170
+ if self.memory and len(self.memory.buffer_as_messages) > 0:
171
+ relevant_documents.append(
172
+ Document(
173
+ page_content="""Following, the previous question and answers. Use these information only when in the question there are unspecified references:\n{}\n\n""".format(
174
+ self.memory.buffer_as_str))
175
+ )
176
  return relevant_documents
177
 
178
  def get_all_context_by_document(self, doc_id):
 
187
  relevant_documents = multi_query_retriever.get_relevant_documents(query)
188
  return relevant_documents
189
 
190
+ def get_text_from_document(self, pdf_file_path, chunk_size=-1, perc_overlap=0.1, include=(), verbose=False):
191
+ """
192
+ Extract text from documents using Grobid, if chunk_size is < 0 it keeps each paragraph separately
193
+ """
194
  if verbose:
195
  print("File", pdf_file_path)
196
  filename = Path(pdf_file_path).stem
 
205
  texts = []
206
  metadatas = []
207
  ids = []
208
+
209
  if chunk_size < 0:
210
  for passage in structure['passages']:
211
  biblio_copy = copy.copy(biblio)
 
229
  metadatas = [biblio for _ in range(len(texts))]
230
  ids = [id for id, t in enumerate(texts)]
231
 
232
+ if "biblio" in include:
233
+ biblio_metadata = copy.copy(biblio)
234
+ biblio_metadata['type'] = "biblio"
235
+ biblio_metadata['section'] = "header"
236
+ for key in ['title', 'authors', 'publication_year']:
237
+ if key in biblio_metadata:
238
+ texts.append("{}: {}".format(key, biblio_metadata[key]))
239
+ metadatas.append(biblio_metadata)
240
+ ids.append(key)
241
+
242
  return texts, metadatas, ids
243
 
244
+ def create_memory_embeddings(self, pdf_path, doc_id=None, chunk_size=500, perc_overlap=0.1, include_biblio=False):
245
+ include = ["biblio"] if include_biblio else []
246
+ texts, metadata, ids = self.get_text_from_document(
247
+ pdf_path,
248
+ chunk_size=chunk_size,
249
+ perc_overlap=perc_overlap,
250
+ include=include)
251
  if doc_id:
252
  hash = doc_id
253
  else:
254
  hash = metadata[0]['hash']
255
 
256
  if hash not in self.embeddings_dict.keys():
257
+ self.embeddings_dict[hash] = Chroma.from_texts(texts,
258
+ embedding=self.embedding_function,
259
+ metadatas=metadata,
260
  collection_name=hash)
261
  else:
262
+ # if 'documents' in self.embeddings_dict[hash].get() and len(self.embeddings_dict[hash].get()['documents']) == 0:
263
+ # self.embeddings_dict[hash].delete(ids=self.embeddings_dict[hash].get()['ids'])
264
+ self.embeddings_dict[hash].delete_collection()
265
+ self.embeddings_dict[hash] = Chroma.from_texts(texts,
266
+ embedding=self.embedding_function,
267
+ metadatas=metadata,
268
  collection_name=hash)
269
 
270
  self.embeddings_root_path = None
271
 
272
  return hash
273
 
274
+ def create_embeddings(self, pdfs_dir_path: Path, chunk_size=500, perc_overlap=0.1, include_biblio=False):
275
  input_files = []
276
  for root, dirs, files in os.walk(pdfs_dir_path, followlinks=False):
277
  for file_ in files:
 
288
  if os.path.exists(data_path):
289
  print(data_path, "exists. Skipping it ")
290
  continue
291
+ include = ["biblio"] if include_biblio else []
292
+ texts, metadata, ids = self.get_text_from_document(
293
+ input_file,
294
+ chunk_size=chunk_size,
295
+ perc_overlap=perc_overlap,
296
+ include=include)
297
  filename = metadata[0]['filename']
298
 
299
  vector_db_document = Chroma.from_texts(texts,
document_qa/grobid_processors.py CHANGED
@@ -171,7 +171,7 @@ class GrobidProcessor(BaseProcessor):
171
  }
172
  try:
173
  year = dateparser.parse(doc_biblio.header.date).year
174
- biblio["year"] = year
175
  except:
176
  pass
177
 
 
171
  }
172
  try:
173
  year = dateparser.parse(doc_biblio.header.date).year
174
+ biblio["publication_year"] = year
175
  except:
176
  pass
177
 
pyproject.toml CHANGED
@@ -3,7 +3,7 @@ requires = ["setuptools", "setuptools-scm"]
3
  build-backend = "setuptools.build_meta"
4
 
5
  [tool.bumpversion]
6
- current_version = "0.3.0"
7
  commit = "true"
8
  tag = "true"
9
  tag_name = "v{new_version}"
 
3
  build-backend = "setuptools.build_meta"
4
 
5
  [tool.bumpversion]
6
+ current_version = "0.3.2"
7
  commit = "true"
8
  tag = "true"
9
  tag_name = "v{new_version}"
streamlit_app.py CHANGED
@@ -115,6 +115,7 @@ def clear_memory():
115
 
116
  # @st.cache_resource
117
  def init_qa(model, api_key=None):
 
118
  if model == 'chatgpt-3.5-turbo':
119
  if api_key:
120
  chat = ChatOpenAI(model_name="gpt-3.5-turbo",
@@ -143,7 +144,7 @@ def init_qa(model, api_key=None):
143
  st.stop()
144
  return
145
 
146
- return DocumentQAEngine(chat, embeddings, grobid_url=os.environ['GROBID_URL'])
147
 
148
 
149
  @st.cache_resource
@@ -252,7 +253,8 @@ with st.sidebar:
252
 
253
  st.button(
254
  'Reset chat memory.',
255
- on_click=clear_memory(),
 
256
  help="Clear the conversational memory. Currently implemented to retrain the 4 most recent messages.")
257
 
258
  left_column, right_column = st.columns([1, 1])
@@ -264,7 +266,9 @@ with right_column:
264
  st.markdown(
265
  ":warning: Do not upload sensitive data. We **temporarily** store text from the uploaded PDF documents solely for the purpose of processing your request, and we **do not assume responsibility** for any subsequent use or handling of the data submitted to third parties LLMs.")
266
 
267
- uploaded_file = st.file_uploader("Upload an article", type=("pdf", "txt"), on_change=new_file,
 
 
268
  disabled=st.session_state['model'] is not None and st.session_state['model'] not in
269
  st.session_state['api_keys'],
270
  help="The full-text is extracted using Grobid. ")
@@ -331,7 +335,8 @@ if uploaded_file and not st.session_state.loaded_embeddings:
331
 
332
  st.session_state['doc_id'] = hash = st.session_state['rqa'][model].create_memory_embeddings(tmp_file.name,
333
  chunk_size=chunk_size,
334
- perc_overlap=0.1)
 
335
  st.session_state['loaded_embeddings'] = True
336
  st.session_state.messages = []
337
 
@@ -384,8 +389,7 @@ with right_column:
384
  elif mode == "LLM":
385
  with st.spinner("Generating response..."):
386
  _, text_response = st.session_state['rqa'][model].query_document(question, st.session_state.doc_id,
387
- context_size=context_size,
388
- memory=st.session_state.memory)
389
 
390
  if not text_response:
391
  st.error("Something went wrong. Contact Luca Foppiano (Foppiano.Luca@nims.co.jp) to report the issue.")
@@ -404,11 +408,11 @@ with right_column:
404
  st.write(text_response)
405
  st.session_state.messages.append({"role": "assistant", "mode": mode, "content": text_response})
406
 
407
- for id in range(0, len(st.session_state.messages), 2):
408
- question = st.session_state.messages[id]['content']
409
- if len(st.session_state.messages) > id + 1:
410
- answer = st.session_state.messages[id + 1]['content']
411
- st.session_state.memory.save_context({"input": question}, {"output": answer})
412
 
413
  elif st.session_state.loaded_embeddings and st.session_state.doc_id:
414
  play_old_messages()
 
115
 
116
  # @st.cache_resource
117
  def init_qa(model, api_key=None):
118
+ ## For debug add: callbacks=[PromptLayerCallbackHandler(pl_tags=["langchain", "chatgpt", "document-qa"])])
119
  if model == 'chatgpt-3.5-turbo':
120
  if api_key:
121
  chat = ChatOpenAI(model_name="gpt-3.5-turbo",
 
144
  st.stop()
145
  return
146
 
147
+ return DocumentQAEngine(chat, embeddings, grobid_url=os.environ['GROBID_URL'], memory=st.session_state['memory'])
148
 
149
 
150
  @st.cache_resource
 
253
 
254
  st.button(
255
  'Reset chat memory.',
256
+ key="reset-memory-button",
257
+ on_click=clear_memory,
258
  help="Clear the conversational memory. Currently implemented to retrain the 4 most recent messages.")
259
 
260
  left_column, right_column = st.columns([1, 1])
 
266
  st.markdown(
267
  ":warning: Do not upload sensitive data. We **temporarily** store text from the uploaded PDF documents solely for the purpose of processing your request, and we **do not assume responsibility** for any subsequent use or handling of the data submitted to third parties LLMs.")
268
 
269
+ uploaded_file = st.file_uploader("Upload an article",
270
+ type=("pdf", "txt"),
271
+ on_change=new_file,
272
  disabled=st.session_state['model'] is not None and st.session_state['model'] not in
273
  st.session_state['api_keys'],
274
  help="The full-text is extracted using Grobid. ")
 
335
 
336
  st.session_state['doc_id'] = hash = st.session_state['rqa'][model].create_memory_embeddings(tmp_file.name,
337
  chunk_size=chunk_size,
338
+ perc_overlap=0.1,
339
+ include_biblio=True)
340
  st.session_state['loaded_embeddings'] = True
341
  st.session_state.messages = []
342
 
 
389
  elif mode == "LLM":
390
  with st.spinner("Generating response..."):
391
  _, text_response = st.session_state['rqa'][model].query_document(question, st.session_state.doc_id,
392
+ context_size=context_size)
 
393
 
394
  if not text_response:
395
  st.error("Something went wrong. Contact Luca Foppiano (Foppiano.Luca@nims.co.jp) to report the issue.")
 
408
  st.write(text_response)
409
  st.session_state.messages.append({"role": "assistant", "mode": mode, "content": text_response})
410
 
411
+ # if len(st.session_state.messages) > 1:
412
+ # last_answer = st.session_state.messages[len(st.session_state.messages)-1]
413
+ # if last_answer['role'] == "assistant":
414
+ # last_question = st.session_state.messages[len(st.session_state.messages)-2]
415
+ # st.session_state.memory.save_context({"input": last_question['content']}, {"output": last_answer['content']})
416
 
417
  elif st.session_state.loaded_embeddings and st.session_state.doc_id:
418
  play_old_messages()