Spaces:
Runtime error
Runtime error
clementsan
commited on
Commit
•
9733941
1
Parent(s):
1da1e92
Enable use of document references
Browse files
app.py
CHANGED
@@ -101,6 +101,7 @@ def initialize_llmchain(llm_model, temperature, max_tokens, top_k, vector_db, pr
|
|
101 |
progress(0.75, desc="Defining buffer memory...")
|
102 |
memory = ConversationBufferMemory(
|
103 |
memory_key="chat_history",
|
|
|
104 |
return_messages=True
|
105 |
)
|
106 |
# retriever=vector_db.as_retriever(search_type="similarity", search_kwargs={'k': 3})
|
@@ -113,7 +114,7 @@ def initialize_llmchain(llm_model, temperature, max_tokens, top_k, vector_db, pr
|
|
113 |
chain_type="stuff",
|
114 |
memory=memory,
|
115 |
# combine_docs_chain_kwargs={"prompt": your_prompt})
|
116 |
-
|
117 |
# return_generated_question=True,
|
118 |
# verbose=True,
|
119 |
)
|
@@ -162,11 +163,20 @@ def conversation(message, history):
|
|
162 |
|
163 |
# Generate response using QA chain
|
164 |
response = qa_chain({"question": message, "chat_history": formatted_chat_history})
|
165 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
166 |
|
167 |
# Append user message and response to chat history
|
168 |
-
new_history = history + [(message,
|
169 |
-
return gr.update(value=""), new_history
|
|
|
170 |
|
171 |
|
172 |
def upload_file(file_obj):
|
@@ -188,7 +198,7 @@ def demo():
|
|
188 |
"""<center><h2>PDF-based chatbot (powered by LangChain and open-source LLMs)</center></h2>
|
189 |
<h3>Ask any questions about your PDF documents, along with follow-ups</h3>
|
190 |
<b>Note:</b> This AI assistant performs retrieval-augmented generation from your PDF documents. \
|
191 |
-
When generating answers, it takes past questions into account (via conversational memory), and
|
192 |
<br><b>Warning:</b> This space uses the free CPU Basic hardware from Hugging Face. Some steps and LLM models used below (free inference endpoints) can take some time to generate a reply.
|
193 |
""")
|
194 |
with gr.Tab("Step 1 - Document pre-processing"):
|
@@ -199,7 +209,7 @@ def demo():
|
|
199 |
db_btn = gr.Radio(["ChromaDB"], label="Vector database type", value = "ChromaDB", type="index", info="Choose your vector database")
|
200 |
with gr.Accordion("Advanced options - Document text splitter", open=False):
|
201 |
with gr.Row():
|
202 |
-
slider_chunk_size = gr.Slider(minimum = 100, maximum = 1000, value=
|
203 |
with gr.Row():
|
204 |
slider_chunk_overlap = gr.Slider(minimum = 10, maximum = 200, value=40, step=10, label="Chunk overlap", info="Chunk overlap", interactive=True)
|
205 |
with gr.Row():
|
@@ -222,6 +232,13 @@ def demo():
|
|
222 |
|
223 |
with gr.Tab("Step 3 - Conversation with chatbot"):
|
224 |
chatbot = gr.Chatbot(height=300)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
225 |
with gr.Row():
|
226 |
msg = gr.Textbox(placeholder="Type message", container=True)
|
227 |
with gr.Row():
|
@@ -230,13 +247,29 @@ def demo():
|
|
230 |
|
231 |
# Preprocessing events
|
232 |
#upload_btn.upload(upload_file, inputs=[upload_btn], outputs=[document])
|
233 |
-
db_btn.click(initialize_database,
|
234 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
235 |
|
236 |
# Chatbot events
|
237 |
-
msg.submit(conversation,
|
238 |
-
|
239 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
240 |
demo.queue().launch(debug=True)
|
241 |
|
242 |
|
|
|
101 |
progress(0.75, desc="Defining buffer memory...")
|
102 |
memory = ConversationBufferMemory(
|
103 |
memory_key="chat_history",
|
104 |
+
output_key='answer',
|
105 |
return_messages=True
|
106 |
)
|
107 |
# retriever=vector_db.as_retriever(search_type="similarity", search_kwargs={'k': 3})
|
|
|
114 |
chain_type="stuff",
|
115 |
memory=memory,
|
116 |
# combine_docs_chain_kwargs={"prompt": your_prompt})
|
117 |
+
return_source_documents=True,
|
118 |
# return_generated_question=True,
|
119 |
# verbose=True,
|
120 |
)
|
|
|
163 |
|
164 |
# Generate response using QA chain
|
165 |
response = qa_chain({"question": message, "chat_history": formatted_chat_history})
|
166 |
+
response_answer = response["answer"]
|
167 |
+
response_sources = response["source_documents"]
|
168 |
+
response_source1 = response_sources[0].page_content.strip()
|
169 |
+
response_source2 = response_sources[1].page_content.strip()
|
170 |
+
# Langchain sources are zero-based
|
171 |
+
response_source1_page = response_sources[0].metadata["page"] + 1
|
172 |
+
response_source2_page = response_sources[1].metadata["page"] + 1
|
173 |
+
# print ('chat response: ', response_answer)
|
174 |
+
# print('DB source', response_sources)
|
175 |
|
176 |
# Append user message and response to chat history
|
177 |
+
new_history = history + [(message, response_answer)]
|
178 |
+
# return gr.update(value=""), new_history, response_sources[0], response_sources[1]
|
179 |
+
return gr.update(value=""), new_history, response_source1, response_source1_page, response_source2, response_source2_page
|
180 |
|
181 |
|
182 |
def upload_file(file_obj):
|
|
|
198 |
"""<center><h2>PDF-based chatbot (powered by LangChain and open-source LLMs)</center></h2>
|
199 |
<h3>Ask any questions about your PDF documents, along with follow-ups</h3>
|
200 |
<b>Note:</b> This AI assistant performs retrieval-augmented generation from your PDF documents. \
|
201 |
+
When generating answers, it takes past questions into account (via conversational memory), and includes document references for clarity purposes.</i>
|
202 |
<br><b>Warning:</b> This space uses the free CPU Basic hardware from Hugging Face. Some steps and LLM models used below (free inference endpoints) can take some time to generate a reply.
|
203 |
""")
|
204 |
with gr.Tab("Step 1 - Document pre-processing"):
|
|
|
209 |
db_btn = gr.Radio(["ChromaDB"], label="Vector database type", value = "ChromaDB", type="index", info="Choose your vector database")
|
210 |
with gr.Accordion("Advanced options - Document text splitter", open=False):
|
211 |
with gr.Row():
|
212 |
+
slider_chunk_size = gr.Slider(minimum = 100, maximum = 1000, value=600, step=20, label="Chunk size", info="Chunk size", interactive=True)
|
213 |
with gr.Row():
|
214 |
slider_chunk_overlap = gr.Slider(minimum = 10, maximum = 200, value=40, step=10, label="Chunk overlap", info="Chunk overlap", interactive=True)
|
215 |
with gr.Row():
|
|
|
232 |
|
233 |
with gr.Tab("Step 3 - Conversation with chatbot"):
|
234 |
chatbot = gr.Chatbot(height=300)
|
235 |
+
with gr.Accordion("Advanced - Document references", open=False):
|
236 |
+
with gr.Row():
|
237 |
+
doc_source1 = gr.Textbox(label="Reference 1", lines=2, container=True, scale=20)
|
238 |
+
source1_page = gr.Number(label="Page", scale=1)
|
239 |
+
with gr.Row():
|
240 |
+
doc_source2 = gr.Textbox(label="Reference 2", lines=2, container=True, scale=20)
|
241 |
+
source2_page = gr.Number(label="Page", scale=1)
|
242 |
with gr.Row():
|
243 |
msg = gr.Textbox(placeholder="Type message", container=True)
|
244 |
with gr.Row():
|
|
|
247 |
|
248 |
# Preprocessing events
|
249 |
#upload_btn.upload(upload_file, inputs=[upload_btn], outputs=[document])
|
250 |
+
db_btn.click(initialize_database, \
|
251 |
+
inputs=[document, slider_chunk_size, slider_chunk_overlap], \
|
252 |
+
outputs=[vector_db, db_progress])
|
253 |
+
qachain_btn.click(initialize_LLM, \
|
254 |
+
inputs=[llm_btn, slider_temperature, slider_maxtokens, slider_topk, vector_db], \
|
255 |
+
outputs=[llm_progress]).then(lambda:[None,"",0,"",0], \
|
256 |
+
inputs=None, \
|
257 |
+
outputs=[chatbot, doc_source1, source1_page, doc_source2, source2_page], \
|
258 |
+
queue=False)
|
259 |
|
260 |
# Chatbot events
|
261 |
+
msg.submit(conversation, \
|
262 |
+
inputs=[msg, chatbot], \
|
263 |
+
outputs=[msg, chatbot, doc_source1, source1_page, doc_source2, source2_page], \
|
264 |
+
queue=False)
|
265 |
+
submit_btn.click(conversation, \
|
266 |
+
inputs=[msg, chatbot], \
|
267 |
+
outputs=[msg, chatbot, doc_source1, source1_page, doc_source2, source2_page], \
|
268 |
+
queue=False)
|
269 |
+
clear_btn.click(lambda:[None,"",0,"",0], \
|
270 |
+
inputs=None, \
|
271 |
+
outputs=[chatbot, doc_source1, source1_page, doc_source2, source2_page], \
|
272 |
+
queue=False)
|
273 |
demo.queue().launch(debug=True)
|
274 |
|
275 |
|