|
a |
|
b/app/frontend/streamlit_app_talk2knowledgegraphs.py |
|
|
1 |
#!/usr/bin/env python3 |
|
|
2 |
|
|
|
3 |
""" |
|
|
4 |
Talk2KnowledgeGraphs: A Streamlit app for the Talk2KnowledgeGraphs graph. |
|
|
5 |
""" |
|
|
6 |
|
|
|
7 |
import os |
|
|
8 |
import sys |
|
|
9 |
import random |
|
|
10 |
import streamlit as st |
|
|
11 |
import pandas as pd |
|
|
12 |
import hydra |
|
|
13 |
from streamlit_feedback import streamlit_feedback |
|
|
14 |
from langchain_core.messages import SystemMessage, HumanMessage, AIMessage |
|
|
15 |
from langchain_core.messages import ChatMessage |
|
|
16 |
from langchain_core.tracers.context import collect_runs |
|
|
17 |
from langchain.callbacks.tracers import LangChainTracer |
|
|
18 |
from langchain_openai import OpenAIEmbeddings, ChatOpenAI |
|
|
19 |
from langchain_ollama import OllamaEmbeddings, ChatOllama |
|
|
20 |
from utils import streamlit_utils |
|
|
21 |
|
|
|
22 |
sys.path.append("./") |
|
|
23 |
from aiagents4pharma.talk2knowledgegraphs.agents.t2kg_agent import get_app |
|
|
24 |
# from talk2knowledgegraphs.agents.t2kg_agent import get_app |
|
|
25 |
|
|
|
26 |
st.set_page_config( |
|
|
27 |
page_title="Talk2KnowledgeGraphs", |
|
|
28 |
page_icon="đ¤", |
|
|
29 |
layout="wide", |
|
|
30 |
initial_sidebar_state="collapsed", |
|
|
31 |
) |
|
|
32 |
|
|
|
33 |
# Initialize configuration |
|
|
34 |
hydra.core.global_hydra.GlobalHydra.instance().clear() |
|
|
35 |
if "config" not in st.session_state: |
|
|
36 |
# Load Hydra configuration |
|
|
37 |
with hydra.initialize( |
|
|
38 |
version_base=None, |
|
|
39 |
config_path="../../aiagents4pharma/talk2knowledgegraphs/configs", |
|
|
40 |
): |
|
|
41 |
cfg = hydra.compose(config_name="config", overrides=["app/frontend=default"]) |
|
|
42 |
cfg = cfg.app.frontend |
|
|
43 |
st.session_state.config = cfg |
|
|
44 |
else: |
|
|
45 |
cfg = st.session_state.config |
|
|
46 |
|
|
|
47 |
|
|
|
48 |
# st.logo( |
|
|
49 |
# image='docs/VPE.png', |
|
|
50 |
# size='large', |
|
|
51 |
# link='https://github.com/VirtualPatientEngine' |
|
|
52 |
# ) |
|
|
53 |
|
|
|
54 |
# Check if env variable OPENAI_API_KEY exists |
|
|
55 |
if "OPENAI_API_KEY" not in os.environ: |
|
|
56 |
st.error( |
|
|
57 |
"Please set the OPENAI_API_KEY environment \ |
|
|
58 |
variable in the terminal where you run the app." |
|
|
59 |
) |
|
|
60 |
st.stop() |
|
|
61 |
|
|
|
62 |
# Initialize current user |
|
|
63 |
if "current_user" not in st.session_state: |
|
|
64 |
st.session_state.current_user = cfg.default_user |
|
|
65 |
|
|
|
66 |
# Initialize chat history |
|
|
67 |
if "messages" not in st.session_state: |
|
|
68 |
st.session_state.messages = [] |
|
|
69 |
|
|
|
70 |
# Initialize session state for SBML file uploader |
|
|
71 |
# if "sbml_key" not in st.session_state: |
|
|
72 |
# st.session_state.sbml_key = 0 |
|
|
73 |
|
|
|
74 |
# Initialize session state for pre-clinical data package uploader |
|
|
75 |
if "data_package_key" not in st.session_state: |
|
|
76 |
st.session_state.data_package_key = 0 |
|
|
77 |
|
|
|
78 |
# Initialize session state for patient gene expression data uploader |
|
|
79 |
if "endotype_key" not in st.session_state: |
|
|
80 |
st.session_state.endotype_key = 0 |
|
|
81 |
|
|
|
82 |
# Initialize session state for uploaded files |
|
|
83 |
if "uploaded_files" not in st.session_state: |
|
|
84 |
st.session_state.uploaded_files = [] |
|
|
85 |
|
|
|
86 |
# Make directories if not exists |
|
|
87 |
os.makedirs(cfg.upload_data_dir, exist_ok=True) |
|
|
88 |
|
|
|
89 |
# Initialize project_name for Langsmith |
|
|
90 |
if "project_name" not in st.session_state: |
|
|
91 |
# st.session_state.project_name = str(st.session_state.user_name) + '@' + str(uuid.uuid4()) |
|
|
92 |
st.session_state.project_name = "T2KG-" + str(random.randint(1000, 9999)) |
|
|
93 |
|
|
|
94 |
# Initialize run_id for Langsmith |
|
|
95 |
if "run_id" not in st.session_state: |
|
|
96 |
st.session_state.run_id = None |
|
|
97 |
|
|
|
98 |
# Initialize graph |
|
|
99 |
if "unique_id" not in st.session_state: |
|
|
100 |
st.session_state.unique_id = random.randint(1, 1000) |
|
|
101 |
|
|
|
102 |
# Initialize the LLM model |
|
|
103 |
if "llm_model" not in st.session_state: |
|
|
104 |
st.session_state.llm_model = tuple(cfg.openai_llms + cfg.ollama_llms)[0] |
|
|
105 |
|
|
|
106 |
# Initialize the app with default LLM model for the first time |
|
|
107 |
if "app" not in st.session_state: |
|
|
108 |
# Initialize the app |
|
|
109 |
if st.session_state.llm_model in cfg.openai_llms: |
|
|
110 |
print("Using OpenAI LLM model") |
|
|
111 |
st.session_state.app = get_app( |
|
|
112 |
st.session_state.unique_id, |
|
|
113 |
llm_model=ChatOpenAI( |
|
|
114 |
model=st.session_state.llm_model, temperature=cfg.temperature |
|
|
115 |
), |
|
|
116 |
) |
|
|
117 |
else: |
|
|
118 |
print("Using Ollama LLM model") |
|
|
119 |
st.session_state.app = get_app( |
|
|
120 |
st.session_state.unique_id, |
|
|
121 |
llm_model=ChatOllama( |
|
|
122 |
model=st.session_state.llm_model, temperature=cfg.temperature |
|
|
123 |
), |
|
|
124 |
) |
|
|
125 |
|
|
|
126 |
if "topk_nodes" not in st.session_state: |
|
|
127 |
# Subgraph extraction settings |
|
|
128 |
st.session_state.topk_nodes = cfg.reasoning_subgraph_topk_nodes |
|
|
129 |
st.session_state.topk_edges = cfg.reasoning_subgraph_topk_edges |
|
|
130 |
|
|
|
131 |
# Get the app |
|
|
132 |
app = st.session_state.app |
|
|
133 |
|
|
|
134 |
# Apply custom CSS |
|
|
135 |
streamlit_utils.apply_css() |
|
|
136 |
|
|
|
137 |
# Sidebar |
|
|
138 |
with st.sidebar: |
|
|
139 |
st.markdown("**âī¸ Subgraph Extraction Settings**") |
|
|
140 |
topk_nodes = st.slider( |
|
|
141 |
"Top-K (Nodes)", |
|
|
142 |
cfg.reasoning_subgraph_topk_nodes_min, |
|
|
143 |
cfg.reasoning_subgraph_topk_nodes_max, |
|
|
144 |
st.session_state.topk_nodes, |
|
|
145 |
key="st_slider_topk_nodes", |
|
|
146 |
) |
|
|
147 |
st.session_state.topk_nodes = topk_nodes |
|
|
148 |
topk_edges = st.slider( |
|
|
149 |
"Top-K (Edges)", |
|
|
150 |
cfg.reasoning_subgraph_topk_nodes_min, |
|
|
151 |
cfg.reasoning_subgraph_topk_nodes_max, |
|
|
152 |
st.session_state.topk_edges, |
|
|
153 |
key="st_slider_topk_edges", |
|
|
154 |
) |
|
|
155 |
st.session_state.topk_edges = topk_edges |
|
|
156 |
|
|
|
157 |
# Main layout of the app split into two columns |
|
|
158 |
main_col1, main_col2 = st.columns([3, 7]) |
|
|
159 |
# First column |
|
|
160 |
with main_col1: |
|
|
161 |
with st.container(border=True): |
|
|
162 |
# Title |
|
|
163 |
st.write( |
|
|
164 |
""" |
|
|
165 |
<h3 style='margin: 0px; padding-bottom: 10px; font-weight: bold;'> |
|
|
166 |
đ¤ Talk2KnowledgeGraphs |
|
|
167 |
</h3> |
|
|
168 |
""", |
|
|
169 |
unsafe_allow_html=True, |
|
|
170 |
) |
|
|
171 |
|
|
|
172 |
# LLM panel (Only at the front-end for now) |
|
|
173 |
# llms = ["gpt-4o-mini", "gpt-4-turbo", "gpt-3.5-turbo"] |
|
|
174 |
llms = tuple(cfg.openai_llms + cfg.ollama_llms) |
|
|
175 |
st.selectbox( |
|
|
176 |
"Pick an LLM to power the agent", |
|
|
177 |
llms, |
|
|
178 |
index=0, |
|
|
179 |
key="llm_model", |
|
|
180 |
on_change=streamlit_utils.update_llm_model, |
|
|
181 |
) |
|
|
182 |
|
|
|
183 |
# Upload files |
|
|
184 |
streamlit_utils.get_uploaded_files(cfg) |
|
|
185 |
|
|
|
186 |
# Help text |
|
|
187 |
# st.button("Know more â", |
|
|
188 |
# # icon="âšī¸", |
|
|
189 |
# on_click=streamlit_utils.help_button, |
|
|
190 |
# use_container_width=False) |
|
|
191 |
|
|
|
192 |
with st.container(border=False, height=500): |
|
|
193 |
prompt = st.chat_input("Say something ...", key="st_chat_input") |
|
|
194 |
|
|
|
195 |
# Second column |
|
|
196 |
with main_col2: |
|
|
197 |
# Chat history panel |
|
|
198 |
with st.container(border=True, height=575): |
|
|
199 |
st.write("#### đŦ Chat History") |
|
|
200 |
|
|
|
201 |
# Display chat messages |
|
|
202 |
for count, message in enumerate(st.session_state.messages): |
|
|
203 |
if message["type"] == "message": |
|
|
204 |
with st.chat_message( |
|
|
205 |
message["content"].role, |
|
|
206 |
avatar="đ¤" if message["content"].role != "user" else "đŠđģâđģ", |
|
|
207 |
): |
|
|
208 |
st.markdown(message["content"].content) |
|
|
209 |
st.empty() |
|
|
210 |
elif message["type"] == "plotly": |
|
|
211 |
streamlit_utils.render_plotly( |
|
|
212 |
message["content"], |
|
|
213 |
key=message["key"], |
|
|
214 |
title=message["title"], |
|
|
215 |
# tool_name=message["tool_name"], |
|
|
216 |
save_chart=False, |
|
|
217 |
) |
|
|
218 |
st.empty() |
|
|
219 |
elif message["type"] == "toggle": |
|
|
220 |
streamlit_utils.render_toggle( |
|
|
221 |
key=message["key"], |
|
|
222 |
toggle_text=message["content"], |
|
|
223 |
toggle_state=message["toggle_state"], |
|
|
224 |
save_toggle=False, |
|
|
225 |
) |
|
|
226 |
st.empty() |
|
|
227 |
elif message["type"] == "dataframe": |
|
|
228 |
streamlit_utils.render_table( |
|
|
229 |
message["content"], |
|
|
230 |
key=message["key"], |
|
|
231 |
# tool_name=message["tool_name"], |
|
|
232 |
save_table=False, |
|
|
233 |
) |
|
|
234 |
st.empty() |
|
|
235 |
elif message["type"] == "graph": |
|
|
236 |
streamlit_utils.render_graph( |
|
|
237 |
message["content"], key=message["key"], save_graph=False |
|
|
238 |
) |
|
|
239 |
st.empty() |
|
|
240 |
|
|
|
241 |
# When the user asks a question |
|
|
242 |
if prompt: |
|
|
243 |
# Display user prompt |
|
|
244 |
prompt_msg = ChatMessage(prompt, role="user") |
|
|
245 |
st.session_state.messages.append({"type": "message", "content": prompt_msg}) |
|
|
246 |
with st.chat_message("user", avatar="đŠđģâđģ"): |
|
|
247 |
st.markdown(prompt) |
|
|
248 |
st.empty() |
|
|
249 |
|
|
|
250 |
# Auxiliary visualization-related variables |
|
|
251 |
graphs_visuals = [] |
|
|
252 |
with st.chat_message("assistant", avatar="đ¤"): |
|
|
253 |
# with st.spinner("Fetching response ..."): |
|
|
254 |
with st.spinner(): |
|
|
255 |
# Get chat history |
|
|
256 |
history = [ |
|
|
257 |
(m["content"].role, m["content"].content) |
|
|
258 |
for m in st.session_state.messages |
|
|
259 |
if m["type"] == "message" |
|
|
260 |
] |
|
|
261 |
# Convert chat history to ChatMessage objects |
|
|
262 |
chat_history = [ |
|
|
263 |
SystemMessage(content=m[1]) |
|
|
264 |
if m[0] == "system" |
|
|
265 |
else HumanMessage(content=m[1]) |
|
|
266 |
if m[0] == "human" |
|
|
267 |
else AIMessage(content=m[1]) |
|
|
268 |
for m in history |
|
|
269 |
] |
|
|
270 |
|
|
|
271 |
# Prepare LLM and embedding model for updating the agent |
|
|
272 |
if st.session_state.llm_model in cfg.openai_llms: |
|
|
273 |
llm_model = ChatOpenAI( |
|
|
274 |
model=st.session_state.llm_model, |
|
|
275 |
temperature=cfg.temperature, |
|
|
276 |
) |
|
|
277 |
emb_model = OpenAIEmbeddings(model=cfg.openai_embeddings[0]) |
|
|
278 |
else: |
|
|
279 |
llm_model = ChatOllama( |
|
|
280 |
model=st.session_state.llm_model, |
|
|
281 |
temperature=cfg.temperature, |
|
|
282 |
) |
|
|
283 |
emb_model = OllamaEmbeddings(model=cfg.ollama_embeddings[0]) |
|
|
284 |
|
|
|
285 |
# Create config for the agent |
|
|
286 |
config = {"configurable": {"thread_id": st.session_state.unique_id}} |
|
|
287 |
app.update_state( |
|
|
288 |
config, |
|
|
289 |
{ |
|
|
290 |
"llm_model": llm_model, |
|
|
291 |
"embedding_model": emb_model, |
|
|
292 |
"uploaded_files": st.session_state.uploaded_files, |
|
|
293 |
"topk_nodes": st.session_state.topk_nodes, |
|
|
294 |
"topk_edges": st.session_state.topk_edges, |
|
|
295 |
"dic_source_graph": [ |
|
|
296 |
{ |
|
|
297 |
"name": st.session_state.config["kg_name"], |
|
|
298 |
"kg_pyg_path": st.session_state.config["kg_pyg_path"], |
|
|
299 |
"kg_text_path": st.session_state.config["kg_text_path"], |
|
|
300 |
} |
|
|
301 |
], |
|
|
302 |
}, |
|
|
303 |
) |
|
|
304 |
|
|
|
305 |
# Update the agent states |
|
|
306 |
current_state = app.get_state(config) |
|
|
307 |
|
|
|
308 |
ERROR_FLAG = False |
|
|
309 |
with collect_runs() as cb: |
|
|
310 |
# Add Langsmith tracer |
|
|
311 |
tracer = LangChainTracer( |
|
|
312 |
project_name=st.session_state.project_name |
|
|
313 |
) |
|
|
314 |
# Get response from the agent |
|
|
315 |
response = app.invoke( |
|
|
316 |
{"messages": [HumanMessage(content=prompt)]}, |
|
|
317 |
config=config | {"callbacks": [tracer]}, |
|
|
318 |
) |
|
|
319 |
st.session_state.run_id = cb.traced_runs[-1].id |
|
|
320 |
current_state = app.get_state(config) |
|
|
321 |
|
|
|
322 |
# Add response to chat history |
|
|
323 |
assistant_msg = ChatMessage( |
|
|
324 |
response["messages"][-1].content, role="assistant" |
|
|
325 |
) |
|
|
326 |
st.session_state.messages.append( |
|
|
327 |
{"type": "message", "content": assistant_msg} |
|
|
328 |
) |
|
|
329 |
# Display the response in the chat |
|
|
330 |
st.markdown(response["messages"][-1].content) |
|
|
331 |
st.empty() |
|
|
332 |
|
|
|
333 |
# Get the current state of the graph |
|
|
334 |
current_state = app.get_state(config) |
|
|
335 |
|
|
|
336 |
# # Get the messages from the current state |
|
|
337 |
# # and reverse the order |
|
|
338 |
reversed_messages = current_state.values["messages"][::-1] |
|
|
339 |
|
|
|
340 |
# Loop through the reversed messages until a |
|
|
341 |
# HumanMessage is found i.e. the last message |
|
|
342 |
# from the user. This is to display the results |
|
|
343 |
# of the tool calls made by the agent since the |
|
|
344 |
# last message from the user. |
|
|
345 |
for msg in reversed_messages: |
|
|
346 |
# print (msg) |
|
|
347 |
# Break the loop if the message is a HumanMessage |
|
|
348 |
# i.e. the last message from the user |
|
|
349 |
if isinstance(msg, HumanMessage): |
|
|
350 |
break |
|
|
351 |
# Skip the message if it is an AIMessage |
|
|
352 |
# i.e. a message from the agent. An agent |
|
|
353 |
# may make multiple tool calls before the |
|
|
354 |
# final response to the user. |
|
|
355 |
if isinstance(msg, AIMessage): |
|
|
356 |
continue |
|
|
357 |
# Work on the message if it is a ToolMessage |
|
|
358 |
# These may contain additional visuals that |
|
|
359 |
# need to be displayed to the user. |
|
|
360 |
# print("ToolMessage", msg) |
|
|
361 |
# Skip the Tool message if it is an error message |
|
|
362 |
if msg.status == "error": |
|
|
363 |
continue |
|
|
364 |
|
|
|
365 |
# Create a unique message id to identify the tool call |
|
|
366 |
# msg.name is the name of the tool |
|
|
367 |
# msg.tool_call_id is the unique id of the tool call |
|
|
368 |
# st.session_state.run_id is the unique id of the run |
|
|
369 |
uniq_msg_id = ( |
|
|
370 |
msg.name |
|
|
371 |
+ "_" |
|
|
372 |
+ msg.tool_call_id |
|
|
373 |
+ "_" |
|
|
374 |
+ str(st.session_state.run_id) |
|
|
375 |
) |
|
|
376 |
if msg.name in ["subgraph_extraction"]: |
|
|
377 |
print( |
|
|
378 |
"-", |
|
|
379 |
len(current_state.values["dic_extracted_graph"]), |
|
|
380 |
"subgraph_extraction", |
|
|
381 |
) |
|
|
382 |
# Add the graph into the visuals list |
|
|
383 |
latest_graph = current_state.values["dic_extracted_graph"][ |
|
|
384 |
-1 |
|
|
385 |
] |
|
|
386 |
if current_state.values["dic_extracted_graph"]: |
|
|
387 |
graphs_visuals.append( |
|
|
388 |
{ |
|
|
389 |
"content": latest_graph["graph_dict"], |
|
|
390 |
"key": "subgraph_" + uniq_msg_id, |
|
|
391 |
} |
|
|
392 |
) |
|
|
393 |
|
|
|
394 |
# Visualize the graph |
|
|
395 |
if len(graphs_visuals) > 0: |
|
|
396 |
for count, graph in enumerate(graphs_visuals): |
|
|
397 |
streamlit_utils.render_graph( |
|
|
398 |
graph_dict=graph["content"], key=graph["key"], save_graph=True |
|
|
399 |
) |
|
|
400 |
|
|
|
401 |
# Collect feedback and display the thumbs feedback |
|
|
402 |
if st.session_state.get("run_id"): |
|
|
403 |
feedback = streamlit_feedback( |
|
|
404 |
feedback_type="thumbs", |
|
|
405 |
optional_text_label="[Optional] Please provide an explanation", |
|
|
406 |
on_submit=streamlit_utils.submit_feedback, |
|
|
407 |
key=f"feedback_{st.session_state.run_id}", |
|
|
408 |
) |